diff --git a/libc/bionic/pthread_mutex.cpp b/libc/bionic/pthread_mutex.cpp
index ed90639..e5f2a31 100644
--- a/libc/bionic/pthread_mutex.cpp
+++ b/libc/bionic/pthread_mutex.cpp
@@ -31,6 +31,7 @@
 #include <errno.h>
 #include <limits.h>
 #include <stdatomic.h>
+#include <stdlib.h>
 #include <string.h>
 #include <sys/cdefs.h>
 #include <sys/mman.h>
@@ -130,8 +131,6 @@
     return 0;
 }
 
-#if defined(__LP64__)
-
 // Priority Inheritance mutex implementation
 struct PIMutex {
   // mutex type, can be 0 (normal), 1 (recursive), 2 (errorcheck), constant during lifetime
@@ -172,12 +171,16 @@
     return EBUSY;
 }
 
-static int PIMutexTimedLock(PIMutex& mutex, const timespec* abs_timeout) {
+// Inlining this function in pthread_mutex_lock() add the cost of stack frame instructions on
+// ARM/ARM64, which increases at most 20 percent overhead. So make it noinline.
+static int  __attribute__((noinline)) PIMutexTimedLock(PIMutex& mutex,
+                                                       const timespec* abs_timeout) {
     int ret = PIMutexTryLock(mutex);
     if (__predict_true(ret == 0)) {
         return 0;
     }
     if (ret == EBUSY) {
+        ScopedTrace trace("Contending for pthread mutex");
         ret = -__futex_pi_lock_ex(&mutex.owner_tid, mutex.shared, true, abs_timeout);
     }
     return ret;
@@ -228,7 +231,97 @@
     }
     return EBUSY;
 }
-#endif  // defined(__LP64__)
+
+#if !defined(__LP64__)
+
+namespace PIMutexAllocator {
+// pthread_mutex_t has only 4 bytes in 32-bit programs, which are not enough to hold PIMutex.
+// So we use malloc to allocate PIMutexes and use 16-bit of pthread_mutex_t as indexes to find
+// the allocated PIMutexes. This allows at most 65536 PI mutexes.
+// When calling operations like pthread_mutex_lock/unlock, the 16-bit index is mapped to the
+// corresponding PIMutex. To make the map operation fast, we use a lockless mapping method:
+//   Once a PIMutex is allocated, all the data used to map index to the PIMutex isn't changed until
+//   it is destroyed.
+// Below are the data structures:
+//   // struct Node contains a PIMutex.
+//   typedef Node NodeArray[256];
+//   typedef NodeArray* NodeArrayP;
+//   NodeArrayP nodes[256];
+//
+// A 16-bit index is mapped to Node as below:
+//   (*nodes[index >> 8])[index & 0xff]
+//
+// Also use a free list to allow O(1) finding recycled PIMutexes.
+
+union Node {
+    PIMutex mutex;
+    int next_free_id;  // If not -1, refer to the next node in the free PIMutex list.
+};
+typedef Node NodeArray[256];
+typedef NodeArray* NodeArrayP;
+
+// lock_ protects below items.
+static Lock lock;
+static NodeArrayP* nodes;
+static int next_to_alloc_id;
+static int first_free_id = -1;  // If not -1, refer to the first node in the free PIMutex list.
+
+static inline __always_inline Node& IdToNode(int id) {
+    return (*nodes[id >> 8])[id & 0xff];
+}
+
+static inline __always_inline PIMutex& IdToPIMutex(int id) {
+    return IdToNode(id).mutex;
+}
+
+static int AllocIdLocked() {
+    if (first_free_id != -1) {
+        int result = first_free_id;
+        first_free_id = IdToNode(result).next_free_id;
+        return result;
+    }
+    if (next_to_alloc_id >= 0x10000) {
+        return -1;
+    }
+    int array_pos = next_to_alloc_id >> 8;
+    int node_pos = next_to_alloc_id & 0xff;
+    if (node_pos == 0) {
+        if (array_pos == 0) {
+            nodes = static_cast<NodeArray**>(calloc(256, sizeof(NodeArray*)));
+            if (nodes == nullptr) {
+                return -1;
+            }
+        }
+        nodes[array_pos] = static_cast<NodeArray*>(malloc(sizeof(NodeArray)));
+        if (nodes[array_pos] == nullptr) {
+            return -1;
+        }
+    }
+    return next_to_alloc_id++;
+}
+
+// If succeed, return an id referring to a PIMutex, otherwise return -1.
+// A valid id is in range [0, 0xffff].
+static int AllocId() {
+    lock.lock();
+    int result = AllocIdLocked();
+    lock.unlock();
+    if (result != -1) {
+        memset(&IdToPIMutex(result), 0, sizeof(PIMutex));
+    }
+    return result;
+}
+
+static void FreeId(int id) {
+    lock.lock();
+    IdToNode(id).next_free_id = first_free_id;
+    first_free_id = id;
+    lock.unlock();
+}
+
+}  // namespace PIMutexAllocator
+
+#endif  // !defined(__LP64__)
 
 
 /* Convenience macro, creates a mask of 'bits' bits that starts from
@@ -324,7 +417,7 @@
 
 // For a PI mutex, it includes below fields:
 //   Atomic(uint16_t) state;
-//   PIMutex pi_mutex;
+//   PIMutex pi_mutex;  // uint16_t pi_mutex_id in 32-bit programs
 //
 //   state holds the following fields:
 //
@@ -332,6 +425,7 @@
 //   15-14   type    mutex type, should be 3
 //
 //   pi_mutex holds the state of a PI mutex.
+//   pi_mutex_id holds an integer to find the state of a PI mutex.
 //
 // For a Non-PI mutex, it includes below fields:
 //   Atomic(uint16_t) state;
@@ -351,20 +445,42 @@
 //   thread id.
 //
 // PI mutexes and Non-PI mutexes are distinguished by checking type field in state.
-struct pthread_mutex_internal_t {
-  _Atomic(uint16_t) state;
 #if defined(__LP64__)
-  uint16_t __pad;
-  union {
-    atomic_int owner_tid;
-    PIMutex pi_mutex;
-  };
-  char __reserved[28];
-#else
-  _Atomic(uint16_t) owner_tid;
-#endif
+struct pthread_mutex_internal_t {
+    _Atomic(uint16_t) state;
+    uint16_t __pad;
+    union {
+        atomic_int owner_tid;
+        PIMutex pi_mutex;
+    };
+    char __reserved[28];
+
+    PIMutex& ToPIMutex() {
+        return pi_mutex;
+    }
+
+    void FreePIMutex() {
+    }
 } __attribute__((aligned(4)));
 
+#else
+struct pthread_mutex_internal_t {
+    _Atomic(uint16_t) state;
+    union {
+        _Atomic(uint16_t) owner_tid;
+        uint16_t pi_mutex_id;
+    };
+
+    PIMutex& ToPIMutex() {
+        return PIMutexAllocator::IdToPIMutex(pi_mutex_id);
+    }
+
+    void FreePIMutex() {
+        PIMutexAllocator::FreeId(pi_mutex_id);
+    }
+} __attribute__((aligned(4)));
+#endif
+
 static_assert(sizeof(pthread_mutex_t) == sizeof(pthread_mutex_internal_t),
               "pthread_mutex_t should actually be pthread_mutex_internal_t in implementation.");
 
@@ -407,13 +523,20 @@
     }
 
     if (((*attr & MUTEXATTR_PROTOCOL_MASK) >> MUTEXATTR_PROTOCOL_SHIFT) == PTHREAD_PRIO_INHERIT) {
-#if defined(__LP64__)
-        atomic_init(&mutex->state, MUTEX_TYPE_BITS_WITH_PI);
-        mutex->pi_mutex.type = *attr & MUTEXATTR_TYPE_MASK;
-        mutex->pi_mutex.shared = (*attr & MUTEXATTR_SHARED_MASK) != 0;
-#else
-        return EINVAL;
+#if !defined(__LP64__)
+        if (state & MUTEX_SHARED_MASK) {
+            return EINVAL;
+        }
+        int id = PIMutexAllocator::AllocId();
+        if (id == -1) {
+            return ENOMEM;
+        }
+        mutex->pi_mutex_id = id;
 #endif
+        atomic_init(&mutex->state, MUTEX_TYPE_BITS_WITH_PI);
+        PIMutex& pi_mutex = mutex->ToPIMutex();
+        pi_mutex.type = *attr & MUTEXATTR_TYPE_MASK;
+        pi_mutex.shared = (*attr & MUTEXATTR_SHARED_MASK) != 0;
     } else {
         atomic_init(&mutex->state, state);
         atomic_init(&mutex->owner_tid, 0);
@@ -668,18 +791,20 @@
 
     uint16_t old_state = atomic_load_explicit(&mutex->state, memory_order_relaxed);
     uint16_t mtype = (old_state & MUTEX_TYPE_MASK);
-    uint16_t shared = (old_state & MUTEX_SHARED_MASK);
     // Avoid slowing down fast path of normal mutex lock operation.
     if (__predict_true(mtype == MUTEX_TYPE_BITS_NORMAL)) {
-      if (__predict_true(NonPI::NormalMutexTryLock(mutex, shared) == 0)) {
-        return 0;
-      }
+        uint16_t shared = (old_state & MUTEX_SHARED_MASK);
+        if (__predict_true(NonPI::NormalMutexTryLock(mutex, shared) == 0)) {
+            return 0;
+        }
+    } else if (mtype == MUTEX_TYPE_BITS_WITH_PI) {
+        PIMutex& m = mutex->ToPIMutex();
+        // Handle common case first.
+        if (__predict_true(PIMutexTryLock(m) == 0)) {
+            return 0;
+        }
+        return PIMutexTimedLock(mutex->ToPIMutex(), nullptr);
     }
-#if defined(__LP64__)
-    if (mtype == MUTEX_TYPE_BITS_WITH_PI) {
-        return PIMutexTimedLock(mutex->pi_mutex, nullptr);
-    }
-#endif
     return NonPI::MutexLockWithTimeout(mutex, false, nullptr);
 }
 
@@ -704,11 +829,9 @@
         NonPI::NormalMutexUnlock(mutex, shared);
         return 0;
     }
-#if defined(__LP64__)
     if (mtype == MUTEX_TYPE_BITS_WITH_PI) {
-        return PIMutexUnlock(mutex->pi_mutex);
+        return PIMutexUnlock(mutex->ToPIMutex());
     }
-#endif
 
     // Do we already own this recursive or error-check mutex?
     pid_t tid = __get_thread()->tid;
@@ -752,11 +875,9 @@
         uint16_t shared = (old_state & MUTEX_SHARED_MASK);
         return NonPI::NormalMutexTryLock(mutex, shared);
     }
-#if defined(__LP64__)
     if (mtype == MUTEX_TYPE_BITS_WITH_PI) {
-        return PIMutexTryLock(mutex->pi_mutex);
+        return PIMutexTryLock(mutex->ToPIMutex());
     }
-#endif
 
     // Do we already own this recursive or error-check mutex?
     pid_t tid = __get_thread()->tid;
@@ -813,23 +934,23 @@
             return 0;
         }
     }
-#if defined(__LP64__)
     if (mtype == MUTEX_TYPE_BITS_WITH_PI) {
-        return PIMutexTimedLock(mutex->pi_mutex, abs_timeout);
+        return PIMutexTimedLock(mutex->ToPIMutex(), abs_timeout);
     }
-#endif
     return NonPI::MutexLockWithTimeout(mutex, true, abs_timeout);
 }
 
 int pthread_mutex_destroy(pthread_mutex_t* mutex_interface) {
     pthread_mutex_internal_t* mutex = __get_internal_mutex(mutex_interface);
     uint16_t old_state = atomic_load_explicit(&mutex->state, memory_order_relaxed);
-#if defined(__LP64__)
     uint16_t mtype  = (old_state & MUTEX_TYPE_MASK);
     if (mtype == MUTEX_TYPE_BITS_WITH_PI) {
-        return PIMutexDestroy(mutex->pi_mutex);
+        int result = PIMutexDestroy(mutex->ToPIMutex());
+        if (result == 0) {
+            mutex->FreePIMutex();
+        }
+        return result;
     }
-#endif
     // Store 0xffff to make the mutex unusable. Although POSIX standard says it is undefined
     // behavior to destroy a locked mutex, we prefer not to change mutex->state in that situation.
     if (MUTEX_STATE_BITS_IS_UNLOCKED(old_state) &&