diff --git a/libc/arch-arm/cortex-a9/bionic/memcpy_base.S b/libc/arch-arm/cortex-a9/bionic/memcpy_base.S
index bb08b94..966b9b3 100644
--- a/libc/arch-arm/cortex-a9/bionic/memcpy_base.S
+++ b/libc/arch-arm/cortex-a9/bionic/memcpy_base.S
@@ -133,8 +133,7 @@
         strbcs      ip, [r0], #1
         strbcs      lr, [r0], #1
 
-        ldmfd       sp!, {r0, lr}
-        bx          lr
+        ldmfd       sp!, {r0, pc}
 END(MEMCPY_BASE)
 
 ENTRY_PRIVATE(MEMCPY_BASE_ALIGNED)
diff --git a/libc/arch-arm/cortex-a9/bionic/memset.S b/libc/arch-arm/cortex-a9/bionic/memset.S
index 8ee6ac2..48ba815 100644
--- a/libc/arch-arm/cortex-a9/bionic/memset.S
+++ b/libc/arch-arm/cortex-a9/bionic/memset.S
@@ -136,8 +136,7 @@
         strbcs      r1, [r0], #1
         strbmi      r1, [r0], #1
         subs        r2, r2, r3
-        popls       {r0, r4-r7, lr}   /* return */
-        bxls        lr
+        popls       {r0, r4-r7, pc}   /* return */
 
         /* align the destination to a cache-line */
         mov         r12, r1
@@ -180,8 +179,7 @@
         strhmi      r1, [r0], #2
         movs        r2, r2, lsl #2
         strbcs      r1, [r0]
-        ldmfd       sp!, {r0, r4-r7, lr}
-        bx          lr
+        ldmfd       sp!, {r0, r4-r7, pc}
 END(__memset_large_copy)
 
         .data
diff --git a/libc/arch-arm/generic/bionic/memcmp.S b/libc/arch-arm/generic/bionic/memcmp.S
index c78dbd4..6643d55 100644
--- a/libc/arch-arm/generic/bionic/memcmp.S
+++ b/libc/arch-arm/generic/bionic/memcmp.S
@@ -221,8 +221,7 @@
         bne         8b
 
 9:      /* restore registers and return */
-        ldmfd       sp!, {r4, lr}
-        bx          lr
+        ldmfd       sp!, {r4, pc}
 
 10:     /* process less than 12 bytes */
         cmp         r2, #0
diff --git a/libc/arch-arm/generic/bionic/memcpy.S b/libc/arch-arm/generic/bionic/memcpy.S
index ea5a399..65cba4c 100644
--- a/libc/arch-arm/generic/bionic/memcpy.S
+++ b/libc/arch-arm/generic/bionic/memcpy.S
@@ -194,8 +194,7 @@
 
         /* we're done! restore everything and return */
 1:      ldmfd       sp!, {r5-r11}
-        ldmfd       sp!, {r0, r4, lr}
-        bx          lr
+        ldmfd       sp!, {r0, r4, pc}
 
         /********************************************************************/
 
@@ -385,8 +384,7 @@
 
         /* we're done! restore sp and spilled registers and return */
         add         sp,  sp, #28
-        ldmfd       sp!, {r0, r4, lr}
-        bx          lr
+        ldmfd       sp!, {r0, r4, pc}
 END(memcpy)
 
         // Only reached when the __memcpy_chk check fails.
diff --git a/libc/arch-arm/generic/bionic/memset.S b/libc/arch-arm/generic/bionic/memset.S
index d17a9c4..b8eabbf 100644
--- a/libc/arch-arm/generic/bionic/memset.S
+++ b/libc/arch-arm/generic/bionic/memset.S
@@ -82,8 +82,7 @@
         strbcs      r1, [r0], #1
         strbmi      r1, [r0], #1
         subs        r2, r2, r3
-        popls       {r0, r4-r7, lr}    /* return */
-        bxls        lr
+        popls       {r0, r4-r7, pc}    /* return */
 
         /* align the destination to a cache-line */
         mov         r12, r1
@@ -126,8 +125,7 @@
         strhmi      r1, [r0], #2
         movs        r2, r2, lsl #2
         strbcs      r1, [r0]
-        ldmfd       sp!, {r0, r4-r7, lr}
-        bx          lr
+        ldmfd       sp!, {r0, r4-r7, pc}
 END(memset)
 
         .data
diff --git a/libc/arch-arm/krait/bionic/memcpy_base.S b/libc/arch-arm/krait/bionic/memcpy_base.S
index 035dcf1..6c098ac 100644
--- a/libc/arch-arm/krait/bionic/memcpy_base.S
+++ b/libc/arch-arm/krait/bionic/memcpy_base.S
@@ -118,6 +118,5 @@
         strbcs      ip, [r0], #1
         strbcs      lr, [r0], #1
 
-        ldmfd       sp!, {r0, lr}
-        bx          lr
+        ldmfd       sp!, {r0, pc}
 END(MEMCPY_BASE)
diff --git a/linker/dlfcn.cpp b/linker/dlfcn.cpp
index ef454ab..4993568 100644
--- a/linker/dlfcn.cpp
+++ b/linker/dlfcn.cpp
@@ -67,9 +67,10 @@
   do_android_update_LD_LIBRARY_PATH(ld_library_path);
 }
 
-static void* dlopen_ext(const char* filename, int flags, const android_dlextinfo* extinfo) {
+static void* dlopen_ext(const char* filename, int flags,
+                        const android_dlextinfo* extinfo, soinfo* caller) {
   ScopedPthreadMutexLocker locker(&g_dl_mutex);
-  soinfo* result = do_dlopen(filename, flags, extinfo);
+  soinfo* result = do_dlopen(filename, flags, extinfo, caller);
   if (result == nullptr) {
     __bionic_format_dlerror("dlopen failed", linker_get_error_buffer());
     return nullptr;
@@ -78,11 +79,15 @@
 }
 
 void* android_dlopen_ext(const char* filename, int flags, const android_dlextinfo* extinfo) {
-  return dlopen_ext(filename, flags, extinfo);
+  void* caller_addr = __builtin_return_address(0);
+  soinfo* caller = find_containing_library(caller_addr);
+  return dlopen_ext(filename, flags, extinfo, caller);
 }
 
 void* dlopen(const char* filename, int flags) {
-  return dlopen_ext(filename, flags, nullptr);
+  void* caller_addr = __builtin_return_address(0);
+  soinfo* caller = find_containing_library(caller_addr);
+  return dlopen_ext(filename, flags, nullptr, caller);
 }
 
 void* dlsym(void* handle, const char* symbol) {
diff --git a/linker/linker.cpp b/linker/linker.cpp
index 611edb3..f71e5e0 100644
--- a/linker/linker.cpp
+++ b/linker/linker.cpp
@@ -1469,9 +1469,17 @@
   return global_group;
 }
 
-static bool find_libraries(soinfo* start_with, const char* const library_names[],
-      size_t library_names_count, soinfo* soinfos[], std::vector<soinfo*>* ld_preloads,
-      size_t ld_preloads_count, int rtld_flags, const android_dlextinfo* extinfo) {
+// add_as_children - add first-level loaded libraries (i.e. library_names[], but
+// not their transitive dependencies) as children of the start_with library.
+// This is false when find_libraries is called for dlopen(), when newly loaded
+// libraries must form a disjoint tree.
+static bool find_libraries(soinfo* start_with,
+                           const char* const library_names[],
+                           size_t library_names_count, soinfo* soinfos[],
+                           std::vector<soinfo*>* ld_preloads,
+                           size_t ld_preloads_count, int rtld_flags,
+                           const android_dlextinfo* extinfo,
+                           bool add_as_children) {
   // Step 0: prepare.
   LoadTaskList load_tasks;
   for (size_t i = 0; i < library_names_count; ++i) {
@@ -1519,7 +1527,7 @@
       return false;
     }
 
-    if (needed_by != nullptr) {
+    if (needed_by != nullptr && (needed_by != start_with || add_as_children)) {
       needed_by->add_child(si);
     }
 
@@ -1546,8 +1554,8 @@
   // Step 2: link libraries.
   soinfo::soinfo_list_t local_group;
   walk_dependencies_tree(
-      start_with == nullptr ? soinfos : &start_with,
-      start_with == nullptr ? soinfos_count : 1,
+      (start_with != nullptr && add_as_children) ? &start_with : soinfos,
+      (start_with != nullptr && add_as_children) ? 1 : soinfos_count,
       [&] (soinfo* si) {
     local_group.push_back(si);
     return true;
@@ -1579,12 +1587,15 @@
   return linked;
 }
 
-static soinfo* find_library(const char* name, int rtld_flags, const android_dlextinfo* extinfo) {
+static soinfo* find_library(const char* name, int rtld_flags,
+                            const android_dlextinfo* extinfo,
+                            soinfo* needed_by) {
   soinfo* si;
 
   if (name == nullptr) {
     si = somain;
-  } else if (!find_libraries(nullptr, &name, 1, &si, nullptr, 0, rtld_flags, extinfo)) {
+  } else if (!find_libraries(needed_by, &name, 1, &si, nullptr, 0, rtld_flags,
+                             extinfo, /* add_as_children */ false)) {
     return nullptr;
   }
 
@@ -1646,7 +1657,7 @@
           TRACE("deprecated (old format of soinfo): %s needs to unload %s",
               si->get_realpath(), library_name);
 
-          soinfo* needed = find_library(library_name, RTLD_NOLOAD, nullptr);
+          soinfo* needed = find_library(library_name, RTLD_NOLOAD, nullptr, nullptr);
           if (needed != nullptr) {
             // Not found: for example if symlink was deleted between dlopen and dlclose
             // Since we cannot really handle errors at this point - print and continue.
@@ -1708,7 +1719,7 @@
   parse_LD_LIBRARY_PATH(ld_library_path);
 }
 
-soinfo* do_dlopen(const char* name, int flags, const android_dlextinfo* extinfo) {
+soinfo* do_dlopen(const char* name, int flags, const android_dlextinfo* extinfo, soinfo *caller) {
   if ((flags & ~(RTLD_NOW|RTLD_LAZY|RTLD_LOCAL|RTLD_GLOBAL|RTLD_NODELETE|RTLD_NOLOAD)) != 0) {
     DL_ERR("invalid flags to dlopen: %x", flags);
     return nullptr;
@@ -1727,7 +1738,7 @@
   }
 
   ProtectedDataGuard guard;
-  soinfo* si = find_library(name, flags, extinfo);
+  soinfo* si = find_library(name, flags, extinfo, caller);
   if (si != nullptr) {
     si->call_constructors();
   }
@@ -3265,7 +3276,8 @@
 
   if (needed_libraries_count > 0 &&
       !find_libraries(si, needed_library_names, needed_libraries_count, nullptr,
-          &g_ld_preloads, ld_preloads_count, RTLD_GLOBAL, nullptr)) {
+                      &g_ld_preloads, ld_preloads_count, RTLD_GLOBAL, nullptr,
+                      /* add_as_children */ true)) {
     __libc_format_fd(2, "CANNOT LINK EXECUTABLE: %s\n", linker_get_error_buffer());
     exit(EXIT_FAILURE);
   } else if (needed_libraries_count == 0) {
diff --git a/linker/linker.h b/linker/linker.h
index 9bca7f3..b1d2566 100644
--- a/linker/linker.h
+++ b/linker/linker.h
@@ -425,7 +425,7 @@
 
 void do_android_get_LD_LIBRARY_PATH(char*, size_t);
 void do_android_update_LD_LIBRARY_PATH(const char* ld_library_path);
-soinfo* do_dlopen(const char* name, int flags, const android_dlextinfo* extinfo);
+soinfo* do_dlopen(const char* name, int flags, const android_dlextinfo* extinfo, soinfo *caller);
 void do_dlclose(soinfo* si);
 
 int do_dl_iterate_phdr(int (*cb)(dl_phdr_info* info, size_t size, void* data), void* data);
diff --git a/tests/dlfcn_test.cpp b/tests/dlfcn_test.cpp
index 92e3287..00cbdc6 100644
--- a/tests/dlfcn_test.cpp
+++ b/tests/dlfcn_test.cpp
@@ -1072,7 +1072,7 @@
   ASSERT_TRUE(fn != nullptr) << dlerror();
 
   void *p = fn();
-  ASSERT_TRUE(p == nullptr);
+  ASSERT_TRUE(p != nullptr);
 
   dlclose(handle);
 }