Merging upstream version 0.0~git20230113.4362b18.

Signed-off-by: Daniel Baumann <daniel.baumann@progress-linux.org>

29 files changed, 295 insertions, 116 deletions

diff --git a/Makefile b/Makefile
index 37551cb..57975a7 100644
--- a/Makefile
+++ b/Makefile
@@ -20,7 +20,7 @@ MALLOC_IMPL ?= dlmalloc
 # yes or no
 BUILD_LIBC_TOP_HALF ?= yes
 # The directory where we will store intermediate artifacts.
-OBJDIR ?= $(CURDIR)/build
+OBJDIR ?= $(CURDIR)/build/$(TARGET_TRIPLE)
 
 # When the length is no larger than this threshold, we consider the
 # overhead of bulk memory opcodes to outweigh the performance benefit,
@@ -32,10 +32,14 @@ BULK_MEMORY_THRESHOLD ?= 32
 # Variables from this point on are not meant to be overridable via the
 # make command-line.
 
-# Set the target variables. Multiarch triples notably omit the vendor field,
-# which happens to be what we do for the main target triple too.
+# Set the default WASI target triple.
 TARGET_TRIPLE = wasm32-wasi
-MULTIARCH_TRIPLE = wasm32-wasi
+
+# Threaded version necessitates a different traget, as objects from different
+# targets can't be mixed together while linking.
+ifeq ($(THREAD_MODEL), posix)
+TARGET_TRIPLE = wasm32-wasi-threads
+endif
 
 # These variables describe the locations of various files and directories in
 # the source tree.
@@ -195,14 +199,22 @@ LIBC_TOP_HALF_MUSL_SOURCES += \
     $(addprefix $(LIBC_TOP_HALF_MUSL_SRC_DIR)/, \
         env/__init_tls.c \
         stdio/__lockfile.c \
+        stdio/flockfile.c \
+        stdio/ftrylockfile.c \
+        stdio/funlockfile.c \
         thread/__lock.c \
         thread/__wait.c \
         thread/__timedwait.c \
         thread/default_attr.c \
         thread/pthread_attr_destroy.c \
+        thread/pthread_attr_get.c \
         thread/pthread_attr_init.c \
         thread/pthread_attr_setstack.c \
+        thread/pthread_attr_setdetachstate.c \
         thread/pthread_attr_setstacksize.c \
+        thread/pthread_barrier_destroy.c \
+        thread/pthread_barrier_init.c \
+        thread/pthread_barrier_wait.c \
         thread/pthread_cleanup_push.c \
         thread/pthread_cond_broadcast.c \
         thread/pthread_cond_destroy.c \
@@ -215,6 +227,8 @@ LIBC_TOP_HALF_MUSL_SOURCES += \
         thread/pthread_condattr_setclock.c \
         thread/pthread_condattr_setpshared.c \
         thread/pthread_create.c \
+        thread/pthread_detach.c \
+        thread/pthread_equal.c \
         thread/pthread_getspecific.c \
         thread/pthread_join.c \
         thread/pthread_key_create.c \
@@ -256,6 +270,7 @@ LIBC_TOP_HALF_MUSL_SOURCES += \
         thread/sem_timedwait.c \
         thread/sem_trywait.c \
         thread/sem_wait.c \
+        thread/wasm32/wasi_thread_start.s \
     )
 endif
 
@@ -274,12 +289,13 @@ LIBC_TOP_HALF_HEADERS_PRIVATE = $(LIBC_TOP_HALF_DIR)/headers/private
 LIBC_TOP_HALF_SOURCES = $(LIBC_TOP_HALF_DIR)/sources
 LIBC_TOP_HALF_ALL_SOURCES = \
     $(LIBC_TOP_HALF_MUSL_SOURCES) \
-    $(sort $(shell find $(LIBC_TOP_HALF_SOURCES) -name \*.c))
+    $(sort $(shell find $(LIBC_TOP_HALF_SOURCES) -name \*.[cs]))
 
 # Add any extra flags
 CFLAGS = $(EXTRA_CFLAGS)
 # Set the target.
 CFLAGS += --target=$(TARGET_TRIPLE)
+ASMFLAGS += --target=$(TARGET_TRIPLE)
 # WebAssembly floating-point match doesn't trap.
 # TODO: Add -fno-signaling-nans when the compiler supports it.
 CFLAGS += -fno-trapping-math
@@ -326,10 +342,11 @@ CFLAGS += -isystem "$(SYSROOT_INC)"
 # These variables describe the locations of various files and directories in
 # the build tree.
 objs = $(patsubst $(CURDIR)/%.c,$(OBJDIR)/%.o,$(1))
+asmobjs = $(patsubst $(CURDIR)/%.s,$(OBJDIR)/%.o,$(1))
 DLMALLOC_OBJS = $(call objs,$(DLMALLOC_SOURCES))
 EMMALLOC_OBJS = $(call objs,$(EMMALLOC_SOURCES))
 LIBC_BOTTOM_HALF_ALL_OBJS = $(call objs,$(LIBC_BOTTOM_HALF_ALL_SOURCES))
-LIBC_TOP_HALF_ALL_OBJS = $(call objs,$(LIBC_TOP_HALF_ALL_SOURCES))
+LIBC_TOP_HALF_ALL_OBJS = $(call asmobjs,$(call objs,$(LIBC_TOP_HALF_ALL_SOURCES)))
 ifeq ($(MALLOC_IMPL),dlmalloc)
 LIBC_OBJS += $(DLMALLOC_OBJS)
 else ifeq ($(MALLOC_IMPL),emmalloc)
@@ -357,9 +374,9 @@ LIBWASI_EMULATED_SIGNAL_MUSL_OBJS = $(call objs,$(LIBWASI_EMULATED_SIGNAL_MUSL_S
 
 # These variables describe the locations of various files and
 # directories in the generated sysroot tree.
-SYSROOT_LIB := $(SYSROOT)/lib/$(MULTIARCH_TRIPLE)
+SYSROOT_LIB := $(SYSROOT)/lib/$(TARGET_TRIPLE)
 SYSROOT_INC = $(SYSROOT)/include
-SYSROOT_SHARE = $(SYSROOT)/share/$(MULTIARCH_TRIPLE)
+SYSROOT_SHARE = $(SYSROOT)/share/$(TARGET_TRIPLE)
 
 # Files from musl's include directory that we don't want to install in the
 # sysroot's include directory.
@@ -504,6 +521,10 @@ $(OBJDIR)/%.o: $(CURDIR)/%.c include_dirs
 	@mkdir -p "$(@D)"
 	$(CC) $(CFLAGS) -MD -MP -o $@ -c $<
 
+$(OBJDIR)/%.o: $(CURDIR)/%.s include_dirs
+	@mkdir -p "$(@D)"
+	$(CC) $(ASMFLAGS) -o $@ -c $<
+
 -include $(shell find $(OBJDIR) -name \*.d)
 
 $(DLMALLOC_OBJS): CFLAGS += \
@@ -645,11 +666,13 @@ check-symbols: startup_files libc
 	@# its builtin include path first, which produces compiler-specific
 	@# output.
 	@#
-	@# TODO: Undefine __FLOAT128__ for now since it's not in clang 8.0.
-	@# TODO: Filter out __FLT16_* for now, as not all versions of clang have these.
 	@# TODO: Filter out __NO_MATH_ERRNO_ and a few __*WIDTH__ that are new to clang 14.
+	@# TODO: Filter out __GCC_HAVE_SYNC_COMPARE_AND_SWAP_* that are new to clang 16.
 	@# TODO: clang defined __FLT_EVAL_METHOD__ until clang 15, so we force-undefine it
 	@# for older versions.
+	@# TODO: Undefine __wasm_mutable_globals__ and __wasm_sign_ext__, that are new to
+	@# clang 16 for -mcpu=generic.
+	@# TODO: As of clang 16, __GNUC_VA_LIST is #defined without a value.
 	$(CC) $(CFLAGS) "$(SYSROOT_SHARE)/include-all.c" \
 	    -isystem $(SYSROOT_INC) \
 	    -std=gnu17 \
@@ -663,22 +686,24 @@ check-symbols: startup_files libc
 	    -U__clang_version__ \
 	    -U__clang_literal_encoding__ \
 	    -U__clang_wide_literal_encoding__ \
+	    -U__wasm_mutable_globals__ \
+	    -U__wasm_sign_ext__ \
 	    -U__GNUC__ \
 	    -U__GNUC_MINOR__ \
 	    -U__GNUC_PATCHLEVEL__ \
 	    -U__VERSION__ \
-	    -U__FLOAT128__ \
 	    -U__NO_MATH_ERRNO__ \
 	    -U__BITINT_MAXWIDTH__ \
 	    -U__FLT_EVAL_METHOD__ -Wno-builtin-macro-redefined \
 	    | sed -e 's/__[[:upper:][:digit:]]*_ATOMIC_\([[:upper:][:digit:]_]*\)_LOCK_FREE/__compiler_ATOMIC_\1_LOCK_FREE/' \
-	    | grep -v '^#define __FLT16_' \
+	    | sed -e 's/__GNUC_VA_LIST $$/__GNUC_VA_LIST 1/' \
 	    | grep -v '^#define __\(BOOL\|INT_\(LEAST\|FAST\)\(8\|16\|32\|64\)\|INT\|LONG\|LLONG\|SHRT\)_WIDTH__' \
+	    | grep -v '^#define __GCC_HAVE_SYNC_COMPARE_AND_SWAP_\(1\|2\|4\|8\)' \
 	    > "$(SYSROOT_SHARE)/predefined-macros.txt"
 
 	# Check that the computed metadata matches the expected metadata.
 	# This ignores whitespace because on Windows the output has CRLF line endings.
-	diff -wur "$(CURDIR)/expected/$(MULTIARCH_TRIPLE)/$(THREAD_MODEL)" "$(SYSROOT_SHARE)"
+	diff -wur "$(CURDIR)/expected/$(TARGET_TRIPLE)" "$(SYSROOT_SHARE)"
 
 install: finish
 	mkdir -p "$(INSTALL_DIR)"
diff --git a/dlmalloc/src/malloc.c b/dlmalloc/src/malloc.c
index 03da739..6e7a15c 100644
--- a/dlmalloc/src/malloc.c
+++ b/dlmalloc/src/malloc.c
@@ -5214,26 +5214,43 @@ static void internal_inspect_all(mstate m,
 /* ------------------ Exported try_init_allocator -------------------- */
 
 /* Symbol marking the end of data, bss and explicit stack, provided by wasm-ld. */
-extern unsigned char __heap_base;
+extern char __heap_base;
+extern char __heap_end __attribute__((__weak__));
 
 /* Initialize the initial state of dlmalloc to be able to use free memory between __heap_base and initial. */
 static void try_init_allocator(void) {
   /* Check that it is a first-time initialization. */
   assert(!is_initialized(gm));
 
-  char *base = (char *)&__heap_base;
-  /* Calls sbrk(0) that returns the initial memory position. */
-  char *init = (char *)CALL_MORECORE(0);
-  int initial_heap_size = init - base;
+  /* Initialize mstate. */
+  ensure_initialization();
+
+  char *base = &__heap_base;
+  // Try to use the linker pseudo-symbol `__heap_end` for the initial size of
+  // the heap, but if that's not defined due to LLVM being too old perhaps then
+  // round up `base` to the nearest `PAGESIZE`. The initial size of linear
+  // memory will be at least the heap base to this page boundary, and it's then
+  // assumed that the initial linear memory image was truncated at that point.
+  // While this reflects the default behavior of `wasm-ld` it is also possible
+  // for users to craft larger linear memories by passing options to extend
+  // beyond this threshold. In this situation the memory will not be used for
+  // dlmalloc.
+  //
+  // Note that `sbrk(0)`, or in dlmalloc-ese `CALL_MORECORE(0)`, is specifically
+  // not used here. That captures the current size of the heap but is only
+  // correct if the we're the first to try to grow the heap. If the heap has
+  // grown elsewhere, such as a different allocator in place, then this would
+  // incorrectly claim such memroy as our own.
+  char *end = &__heap_end;
+  if (end == NULL)
+    end = (char*) page_align((size_t) base);
+  size_t initial_heap_size = end - base;
 
   /* Check that initial heap is long enough to serve a minimal allocation request. */
   if (initial_heap_size <= MIN_CHUNK_SIZE + TOP_FOOT_SIZE + MALLOC_ALIGNMENT) {
     return;
   }
 
-  /* Initialize mstate. */
-  ensure_initialization();
-
   /* Initialize the dlmalloc internal state. */
   gm->least_addr = base;
   gm->seg.base = base;
diff --git a/expected/wasm32-wasi/posix/defined-symbols.txt b/expected/wasm32-wasi-threads/defined-symbols.txt
index 6e565dc..8dba0f1 100644
--- a/expected/wasm32-wasi/posix/defined-symbols.txt
+++ b/expected/wasm32-wasi-threads/defined-symbols.txt
@@ -46,6 +46,7 @@ __do_cleanup_pop
 __do_cleanup_push
 __do_des
 __do_orphaned_stdio_locks
+__dummy_reference
 __duplocale
 __env_rm_add
 __errno_location
@@ -193,7 +194,6 @@ __progname
 __progname_full
 __pthread_cond_timedwait
 __pthread_create
-__pthread_exit
 __pthread_join
 __pthread_key_create
 __pthread_key_delete
@@ -221,6 +221,7 @@ __qsort_r
 __rand48_step
 __random_lockptr
 __reallocarray
+__register_locked_file
 __release_ptc
 __rem_pio2
 __rem_pio2_large
@@ -354,6 +355,7 @@ __wasi_sock_recv
 __wasi_sock_send
 __wasi_sock_shutdown
 __wasi_thread_spawn
+__wasi_thread_start_C
 __wasilibc_access
 __wasilibc_cwd
 __wasilibc_cwd_lock
@@ -649,6 +651,7 @@ fileno
 fileno_unlocked
 finite
 finitef
+flockfile
 floor
 floorf
 floorl
@@ -704,6 +707,8 @@ ftello
 ftello64
 ftime
 ftruncate
+ftrylockfile
+funlockfile
 futimens
 futimesat
 fwide
@@ -972,9 +977,20 @@ program_invocation_short_name
 pselect
 psignal
 pthread_attr_destroy
+pthread_attr_getdetachstate
+pthread_attr_getguardsize
+pthread_attr_getinheritsched
+pthread_attr_getscope
+pthread_attr_getstack
+pthread_attr_getstacksize
 pthread_attr_init
+pthread_attr_setdetachstate
 pthread_attr_setstack
 pthread_attr_setstacksize
+pthread_barrier_destroy
+pthread_barrier_init
+pthread_barrier_wait
+pthread_barrierattr_getpshared
 pthread_cond_broadcast
 pthread_cond_destroy
 pthread_cond_init
@@ -982,11 +998,13 @@ pthread_cond_signal
 pthread_cond_timedwait
 pthread_cond_wait
 pthread_condattr_destroy
+pthread_condattr_getpshared
 pthread_condattr_init
 pthread_condattr_setclock
 pthread_condattr_setpshared
 pthread_create
-pthread_exit
+pthread_detach
+pthread_equal
 pthread_getspecific
 pthread_join
 pthread_key_create
@@ -1000,6 +1018,10 @@ pthread_mutex_timedlock
 pthread_mutex_trylock
 pthread_mutex_unlock
 pthread_mutexattr_destroy
+pthread_mutexattr_getprotocol
+pthread_mutexattr_getpshared
+pthread_mutexattr_getrobust
+pthread_mutexattr_gettype
 pthread_mutexattr_init
 pthread_mutexattr_setprotocol
 pthread_mutexattr_setpshared
@@ -1016,6 +1038,7 @@ pthread_rwlock_trywrlock
 pthread_rwlock_unlock
 pthread_rwlock_wrlock
 pthread_rwlockattr_destroy
+pthread_rwlockattr_getpshared
 pthread_rwlockattr_init
 pthread_rwlockattr_setpshared
 pthread_self
@@ -1211,6 +1234,8 @@ tgamma
 tgammaf
 tgammal
 thrd_current
+thrd_detach
+thrd_equal
 thrd_sleep
 time
 timegm
diff --git a/expected/wasm32-wasi/posix/include-all.c b/expected/wasm32-wasi-threads/include-all.c
index 0b43b07..0b43b07 100644
--- a/expected/wasm32-wasi/posix/include-all.c
+++ b/expected/wasm32-wasi-threads/include-all.c
diff --git a/expected/wasm32-wasi/posix/predefined-macros.txt b/expected/wasm32-wasi-threads/predefined-macros.txt
index ac90920..a9f2982 100644
--- a/expected/wasm32-wasi/posix/predefined-macros.txt
+++ b/expected/wasm32-wasi-threads/predefined-macros.txt
@@ -2489,6 +2489,7 @@
 #define __DEFINED_wctype_t 
 #define __DEFINED_wint_t 
 #define __FINITE_MATH_ONLY__ 0
+#define __FLOAT128__ 1
 #define __FLT(x) (__IS_REAL(x) && sizeof(x) == sizeof(float))
 #define __FLTCX(x) (__IS_CX(x) && sizeof(x) == sizeof(float complex))
 #define __FLT_DECIMAL_DIG__ 9
@@ -3099,8 +3100,6 @@
 #define __wasm__ 1
 #define __wasm_atomics__ 1
 #define __wasm_bulk_memory__ 1
-#define __wasm_mutable_globals__ 1
-#define __wasm_sign_ext__ 1
 #define _tolower(a) ((a)|0x20)
 #define _toupper(a) ((a)&0x5f)
 #define acos(x) __tg_real_complex(acos, (x))
diff --git a/expected/wasm32-wasi/posix/undefined-symbols.txt b/expected/wasm32-wasi-threads/undefined-symbols.txt
index 7def0a9..a685151 100644
--- a/expected/wasm32-wasi/posix/undefined-symbols.txt
+++ b/expected/wasm32-wasi-threads/undefined-symbols.txt
@@ -1,4 +1,5 @@
 __addtf3
+__data_end
 __divtf3
 __eqtf2
 __extenddftf2
@@ -9,6 +10,7 @@ __fixunstfsi
 __floatsitf
 __floatunsitf
 __getf2
+__global_base
 __gttf2
 __heap_base
 __imported_wasi_snapshot_preview1_args_get
diff --git a/expected/wasm32-wasi/single/defined-symbols.txt b/expected/wasm32-wasi/defined-symbols.txt
index a692632..a692632 100644
--- a/expected/wasm32-wasi/single/defined-symbols.txt
+++ b/expected/wasm32-wasi/defined-symbols.txt
diff --git a/expected/wasm32-wasi/single/include-all.c b/expected/wasm32-wasi/include-all.c
index 86297f3..86297f3 100644
--- a/expected/wasm32-wasi/single/include-all.c
+++ b/expected/wasm32-wasi/include-all.c
diff --git a/expected/wasm32-wasi/single/predefined-macros.txt b/expected/wasm32-wasi/predefined-macros.txt
index 1f7b341..234ad3c 100644
--- a/expected/wasm32-wasi/single/predefined-macros.txt
+++ b/expected/wasm32-wasi/predefined-macros.txt
@@ -2123,7 +2123,6 @@
 #define _POSIX_STREAM_MAX 8
 #define _POSIX_SYMLINK_MAX 255
 #define _POSIX_SYMLOOP_MAX 8
-#define _POSIX_THREADS _POSIX_VERSION
 #define _POSIX_THREAD_ATTR_STACKADDR _POSIX_VERSION
 #define _POSIX_THREAD_ATTR_STACKSIZE _POSIX_VERSION
 #define _POSIX_THREAD_CPUTIME _POSIX_VERSION
@@ -2453,6 +2452,7 @@
 #define __DEFINED_wctype_t 
 #define __DEFINED_wint_t 
 #define __FINITE_MATH_ONLY__ 0
+#define __FLOAT128__ 1
 #define __FLT(x) (__IS_REAL(x) && sizeof(x) == sizeof(float))
 #define __FLTCX(x) (__IS_CX(x) && sizeof(x) == sizeof(float complex))
 #define __FLT_DECIMAL_DIG__ 9
diff --git a/expected/wasm32-wasi/single/undefined-symbols.txt b/expected/wasm32-wasi/undefined-symbols.txt
index b9cdb34..b9cdb34 100644
--- a/expected/wasm32-wasi/single/undefined-symbols.txt
+++ b/expected/wasm32-wasi/undefined-symbols.txt
diff --git a/libc-bottom-half/headers/public/__struct_sockaddr.h b/libc-bottom-half/headers/public/__struct_sockaddr.h
index 9891b90..2e14d95 100644
--- a/libc-bottom-half/headers/public/__struct_sockaddr.h
+++ b/libc-bottom-half/headers/public/__struct_sockaddr.h
@@ -1,9 +1,6 @@
 #ifndef __wasilibc___struct_sockaddr_h
 #define __wasilibc___struct_sockaddr_h
 
-#define __need_STDDEF_H_misc
-#include <stddef.h>
-
 #include <__typedef_sa_family_t.h>
 
 struct sockaddr {
diff --git a/libc-bottom-half/headers/public/__struct_sockaddr_in.h b/libc-bottom-half/headers/public/__struct_sockaddr_in.h
index 73dc5c6..5d4ce51 100644
--- a/libc-bottom-half/headers/public/__struct_sockaddr_in.h
+++ b/libc-bottom-half/headers/public/__struct_sockaddr_in.h
@@ -1,15 +1,12 @@
 #ifndef __wasilibc___struct_sockaddr_in_h
 #define __wasilibc___struct_sockaddr_in_h
 
-#define __need_STDDEF_H_misc
-#include <stddef.h>
-
 #include <__typedef_sa_family_t.h>
 #include <__typedef_in_port_t.h>
 #include <__struct_in_addr.h>
 
 struct sockaddr_in {
-    _Alignas(max_align_t) sa_family_t sin_family;
+    __attribute__((aligned(__BIGGEST_ALIGNMENT__))) sa_family_t sin_family;
     in_port_t sin_port;
     struct in_addr sin_addr;
 };
diff --git a/libc-bottom-half/headers/public/__struct_sockaddr_in6.h b/libc-bottom-half/headers/public/__struct_sockaddr_in6.h
index a220f91..98703dc 100644
--- a/libc-bottom-half/headers/public/__struct_sockaddr_in6.h
+++ b/libc-bottom-half/headers/public/__struct_sockaddr_in6.h
@@ -1,15 +1,12 @@
 #ifndef __wasilibc___struct_sockaddr_in6_h
 #define __wasilibc___struct_sockaddr_in6_h
 
-#define __need_STDDEF_H_misc
-#include <stddef.h>
-
 #include <__typedef_sa_family_t.h>
 #include <__typedef_in_port_t.h>
 #include <__struct_in6_addr.h>
 
 struct sockaddr_in6 {
-    _Alignas(max_align_t) sa_family_t sin6_family;
+    __attribute__((aligned(__BIGGEST_ALIGNMENT__))) sa_family_t sin6_family;
     in_port_t sin6_port;
     unsigned sin6_flowinfo;
     struct in6_addr sin6_addr;
diff --git a/libc-bottom-half/headers/public/__struct_sockaddr_storage.h b/libc-bottom-half/headers/public/__struct_sockaddr_storage.h
index 1ae26a7..b4ebad2 100644
--- a/libc-bottom-half/headers/public/__struct_sockaddr_storage.h
+++ b/libc-bottom-half/headers/public/__struct_sockaddr_storage.h
@@ -1,13 +1,10 @@
 #ifndef __wasilibc___struct_sockaddr_storage_h
 #define __wasilibc___struct_sockaddr_storage_h
 
-#define __need_STDDEF_H_misc
-#include <stddef.h>
-
 #include <__typedef_sa_family_t.h>
 
 struct sockaddr_storage {
-    _Alignas(max_align_t) sa_family_t ss_family;
+    __attribute__((aligned(__BIGGEST_ALIGNMENT__))) sa_family_t ss_family;
     char __ss_data[32];
 };
 
diff --git a/libc-bottom-half/headers/public/__struct_sockaddr_un.h b/libc-bottom-half/headers/public/__struct_sockaddr_un.h
index a5be6e5..6371194 100644
--- a/libc-bottom-half/headers/public/__struct_sockaddr_un.h
+++ b/libc-bottom-half/headers/public/__struct_sockaddr_un.h
@@ -1,13 +1,10 @@
 #ifndef __wasilibc___struct_sockaddr_un_h
 #define __wasilibc___struct_sockaddr_un_h
 
-#define __need_STDDEF_H_misc
-#include <stddef.h>
-
 #include <__typedef_sa_family_t.h>
 
 struct sockaddr_un {
-    _Alignas(max_align_t) sa_family_t sun_family;
+    __attribute__((aligned(__BIGGEST_ALIGNMENT__))) sa_family_t sun_family;
 };
 
 #endif
diff --git a/libc-bottom-half/headers/public/wasi/api.h b/libc-bottom-half/headers/public/wasi/api.h
index 1ab7699..45a6506 100644
--- a/libc-bottom-half/headers/public/wasi/api.h
+++ b/libc-bottom-half/headers/public/wasi/api.h
@@ -2099,7 +2099,7 @@ __wasi_errno_t __wasi_sock_shutdown(
  *
  * @see https://github.com/WebAssembly/wasi-threads/#readme
  */
-__wasi_errno_t __wasi_thread_spawn(
+int32_t __wasi_thread_spawn(
     /**
      * A pointer to an opaque struct to be passed to the module's entry
      * function.
diff --git a/libc-bottom-half/sources/__wasilibc_real.c b/libc-bottom-half/sources/__wasilibc_real.c
index 855a2c6..d2e6b71 100644
--- a/libc-bottom-half/sources/__wasilibc_real.c
+++ b/libc-bottom-half/sources/__wasilibc_real.c
@@ -665,8 +665,7 @@ int32_t __imported_wasi_thread_spawn(int32_t arg0) __attribute__((
     __import_name__("thread_spawn")
 ));
 
-__wasi_errno_t __wasi_thread_spawn(void* start_arg) {
-	int32_t ret = __imported_wasi_thread_spawn((int32_t) start_arg);
-    return (uint16_t) ret;
+int32_t __wasi_thread_spawn(void* start_arg) {
+    return __imported_wasi_thread_spawn((int32_t) start_arg);
 }
 #endif
diff --git a/libc-bottom-half/sources/posix.c b/libc-bottom-half/sources/posix.c
index 35dd99b..b3e59ec 100644
--- a/libc-bottom-half/sources/posix.c
+++ b/libc-bottom-half/sources/posix.c
@@ -58,9 +58,9 @@ int __wasilibc_open_nomode(const char *path, int oflag) {
     char *relative_path;
     int dirfd = find_relpath(path, &relative_path);
 
-    // If we can't find a preopen for it, indicate that we lack capabilities.
+    // If we can't find a preopen for it, fail as if we can't find the path.
     if (dirfd == -1) {
-        errno = ENOTCAPABLE;
+        errno = ENOENT;
         return -1;
     }
 
@@ -71,9 +71,9 @@ int access(const char *path, int amode) {
     char *relative_path;
     int dirfd = find_relpath(path, &relative_path);
 
-    // If we can't find a preopen for it, indicate that we lack capabilities.
+    // If we can't find a preopen for it, fail as if we can't find the path.
     if (dirfd == -1) {
-        errno = ENOTCAPABLE;
+        errno = ENOENT;
         return -1;
     }
 
@@ -88,9 +88,9 @@ ssize_t readlink(
     char *relative_path;
     int dirfd = find_relpath(path, &relative_path);
 
-    // If we can't find a preopen for it, indicate that we lack capabilities.
+    // If we can't find a preopen for it, fail as if we can't find the path.
     if (dirfd == -1) {
-        errno = ENOTCAPABLE;
+        errno = ENOENT;
         return -1;
     }
 
@@ -101,9 +101,9 @@ int stat(const char *restrict path, struct stat *restrict buf) {
     char *relative_path;
     int dirfd = find_relpath(path, &relative_path);
 
-    // If we can't find a preopen for it, indicate that we lack capabilities.
+    // If we can't find a preopen for it, fail as if we can't find the path.
     if (dirfd == -1) {
-        errno = ENOTCAPABLE;
+        errno = ENOENT;
         return -1;
     }
 
@@ -114,9 +114,9 @@ int lstat(const char *restrict path, struct stat *restrict buf) {
     char *relative_path;
     int dirfd = find_relpath(path, &relative_path);
 
-    // If we can't find a preopen for it, indicate that we lack capabilities.
+    // If we can't find a preopen for it, fail as if we can't find the path.
     if (dirfd == -1) {
-        errno = ENOTCAPABLE;
+        errno = ENOENT;
         return -1;
     }
 
@@ -127,9 +127,9 @@ int utime(const char *path, const struct utimbuf *times) {
     char *relative_path;
     int dirfd = find_relpath(path, &relative_path);
 
-    // If we can't find a preopen for it, indicate that we lack capabilities.
+    // If we can't find a preopen for it, fail as if we can't find the path.
     if (dirfd == -1) {
-        errno = ENOTCAPABLE;
+        errno = ENOENT;
         return -1;
     }
 
@@ -147,9 +147,9 @@ int utimes(const char *path, const struct timeval times[2]) {
     char *relative_path;
     int dirfd = find_relpath(path, &relative_path);
 
-    // If we can't find a preopen for it, indicate that we lack capabilities.
+    // If we can't find a preopen for it, fail as if we can't find the path.
     if (dirfd == -1) {
-        errno = ENOTCAPABLE;
+        errno = ENOENT;
         return -1;
     }
 
@@ -169,9 +169,9 @@ int unlink(const char *path) {
     char *relative_path;
     int dirfd = find_relpath(path, &relative_path);
 
-    // If we can't find a preopen for it, indicate that we lack capabilities.
+    // If we can't find a preopen for it, fail as if we can't find the path.
     if (dirfd == -1) {
-        errno = ENOTCAPABLE;
+        errno = ENOENT;
         return -1;
     }
 
@@ -185,9 +185,9 @@ int rmdir(const char *path) {
     char *relative_path;
     int dirfd = find_relpath(path, &relative_path);
 
-    // If we can't find a preopen for it, indicate that we lack capabilities.
+    // If we can't find a preopen for it, fail as if we can't find the path.
     if (dirfd == -1) {
-        errno = ENOTCAPABLE;
+        errno = ENOENT;
         return -1;
     }
 
@@ -198,21 +198,21 @@ int remove(const char *path) {
     char *relative_path;
     int dirfd = find_relpath(path, &relative_path);
 
-    // If we can't find a preopen for it, indicate that we lack capabilities.
+    // If we can't find a preopen for it, fail as if we can't find the path.
     if (dirfd == -1) {
-        errno = ENOTCAPABLE;
+        errno = ENOENT;
         return -1;
     }
 
     // First try to remove it as a file.
     int r = __wasilibc_nocwd___wasilibc_unlinkat(dirfd, relative_path);
-    if (r != 0 && (errno == EISDIR || errno == ENOTCAPABLE)) {
+    if (r != 0 && (errno == EISDIR || errno == ENOENT)) {
         // That failed, but it might be a directory.
         r = __wasilibc_nocwd___wasilibc_rmdirat(dirfd, relative_path);
 
         // If it isn't a directory, we lack capabilities to remove it as a file.
         if (errno == ENOTDIR)
-            errno = ENOTCAPABLE;
+            errno = ENOENT;
     }
     return r;
 }
@@ -221,9 +221,9 @@ int mkdir(const char *path, mode_t mode) {
     char *relative_path;
     int dirfd = find_relpath(path, &relative_path);
 
-    // If we can't find a preopen for it, indicate that we lack capabilities.
+    // If we can't find a preopen for it, fail as if we can't find the path.
     if (dirfd == -1) {
-        errno = ENOTCAPABLE;
+        errno = ENOENT;
         return -1;
     }
 
@@ -234,9 +234,9 @@ DIR *opendir(const char *dirname) {
     char *relative_path;
     int dirfd = find_relpath(dirname, &relative_path);
 
-    // If we can't find a preopen for it, indicate that we lack capabilities.
+    // If we can't find a preopen for it, fail as if we can't find the path.
     if (dirfd == -1) {
-        errno = ENOTCAPABLE;
+        errno = ENOENT;
         return NULL;
     }
 
@@ -252,9 +252,9 @@ int scandir(
     char *relative_path;
     int dirfd = find_relpath(dir, &relative_path);
 
-    // If we can't find a preopen for it, indicate that we lack capabilities.
+    // If we can't find a preopen for it, fail as if we can't find the path.
     if (dirfd == -1) {
-        errno = ENOTCAPABLE;
+        errno = ENOENT;
         return -1;
     }
 
@@ -265,9 +265,9 @@ int symlink(const char *target, const char *linkpath) {
     char *relative_path;
     int dirfd = find_relpath(linkpath, &relative_path);
 
-    // If we can't find a preopen for it, indicate that we lack capabilities.
+    // If we can't find a preopen for it, fail as if we can't find the path.
     if (dirfd == -1) {
-        errno = ENOTCAPABLE;
+        errno = ENOENT;
         return -1;
     }
 
@@ -287,8 +287,8 @@ int link(const char *old, const char *new) {
                                            new_dirfd, new_relative_path, 0);
     }
 
-    // We couldn't find a preopen for it; indicate that we lack capabilities.
-    errno = ENOTCAPABLE;
+    // We couldn't find a preopen for it; fail as if we can't find the path.
+    errno = ENOENT;
     return -1;
 }
 
@@ -305,8 +305,8 @@ int rename(const char *old, const char *new) {
                                              new_dirfd, new_relative_path);
     }
 
-    // We couldn't find a preopen for it; indicate that we lack capabilities.
-    errno = ENOTCAPABLE;
+    // We couldn't find a preopen for it; fail as if we can't find the path.
+    errno = ENOENT;
     return -1;
 }
 
@@ -317,9 +317,9 @@ __wasilibc_access(const char *path, int mode, int flags)
     char *relative_path;
     int dirfd = find_relpath(path, &relative_path);
 
-    // If we can't find a preopen for it, indicate that we lack capabilities.
+    // If we can't find a preopen for it, fail as if we can't find the path.
     if (dirfd == -1) {
-        errno = ENOTCAPABLE;
+        errno = ENOENT;
         return -1;
     }
 
@@ -334,9 +334,9 @@ __wasilibc_utimens(const char *path, const struct timespec times[2], int flags)
     char *relative_path;
     int dirfd = find_relpath(path, &relative_path);
 
-    // If we can't find a preopen for it, indicate that we lack capabilities.
+    // If we can't find a preopen for it, fail as if we can't find the path.
     if (dirfd == -1) {
-        errno = ENOTCAPABLE;
+        errno = ENOENT;
         return -1;
     }
 
@@ -351,9 +351,9 @@ __wasilibc_stat(const char *__restrict path, struct stat *__restrict st, int fla
     char *relative_path;
     int dirfd = find_relpath(path, &relative_path);
 
-    // If we can't find a preopen for it, indicate that we lack capabilities.
+    // If we can't find a preopen for it, fail as if we can't find the path.
     if (dirfd == -1) {
-        errno = ENOTCAPABLE;
+        errno = ENOENT;
         return -1;
     }
 
@@ -369,9 +369,9 @@ __wasilibc_link(const char *oldpath, const char *newpath, int flags)
     int old_dirfd = find_relpath(oldpath, &old_relative_path);
     int new_dirfd = find_relpath(newpath, &new_relative_path);
 
-    // If we can't find a preopen for it, indicate that we lack capabilities.
+    // If we can't find a preopen for it, fail as if we can't find the path.
     if (old_dirfd == -1 || new_dirfd == -1) {
-        errno = ENOTCAPABLE;
+        errno = ENOENT;
         return -1;
     }
 
@@ -387,9 +387,9 @@ __wasilibc_link_oldat(int olddirfd, const char *oldpath, const char *newpath, in
     char *new_relative_path;
     int new_dirfd = find_relpath(newpath, &new_relative_path);
 
-    // If we can't find a preopen for it, indicate that we lack capabilities.
+    // If we can't find a preopen for it, fail as if we can't find the path.
     if (new_dirfd == -1) {
-        errno = ENOTCAPABLE;
+        errno = ENOENT;
         return -1;
     }
 
@@ -405,9 +405,9 @@ __wasilibc_link_newat(const char *oldpath, int newdirfd, const char *newpath, in
     char *old_relative_path;
     int old_dirfd = find_relpath(oldpath, &old_relative_path);
 
-    // If we can't find a preopen for it, indicate that we lack capabilities.
+    // If we can't find a preopen for it, fail as if we can't find the path.
     if (old_dirfd == -1) {
-        errno = ENOTCAPABLE;
+        errno = ENOENT;
         return -1;
     }
 
@@ -423,9 +423,9 @@ __wasilibc_rename_oldat(int fromdirfd, const char *from, const char *to)
     char *to_relative_path;
     int to_dirfd = find_relpath(to, &to_relative_path);
 
-    // If we can't find a preopen for it, indicate that we lack capabilities.
+    // If we can't find a preopen for it, fail as if we can't find the path.
     if (to_dirfd == -1) {
-        errno = ENOTCAPABLE;
+        errno = ENOENT;
         return -1;
     }
 
@@ -439,9 +439,9 @@ __wasilibc_rename_newat(const char *from, int todirfd, const char *to)
     char *from_relative_path;
     int from_dirfd = find_relpath(from, &from_relative_path);
 
-    // If we can't find a preopen for it, indicate that we lack capabilities.
+    // If we can't find a preopen for it, fail as if we can't find the path.
     if (from_dirfd == -1) {
-        errno = ENOTCAPABLE;
+        errno = ENOENT;
         return -1;
     }
 
diff --git a/libc-top-half/musl/include/pthread.h b/libc-top-half/musl/include/pthread.h
index 01fe5f2..b14fe82 100644
--- a/libc-top-half/musl/include/pthread.h
+++ b/libc-top-half/musl/include/pthread.h
@@ -85,7 +85,9 @@ extern "C" {
 
 int pthread_create(pthread_t *__restrict, const pthread_attr_t *__restrict, void *(*)(void *), void *__restrict);
 int pthread_detach(pthread_t);
+#ifdef __wasilibc_unmodified_upstream
 _Noreturn void pthread_exit(void *);
+#endif
 int pthread_join(pthread_t, void **);
 
 #ifdef __GNUC__
diff --git a/libc-top-half/musl/include/unistd.h b/libc-top-half/musl/include/unistd.h
index b5cb5c6..0be83e3 100644
--- a/libc-top-half/musl/include/unistd.h
+++ b/libc-top-half/musl/include/unistd.h
@@ -336,7 +336,9 @@ pid_t gettid(void);
 #endif
 #define _POSIX_VDISABLE         0
 
+#if defined(__wasilibc_unmodified_upstream) || defined(_REENTRANT)
 #define _POSIX_THREADS          _POSIX_VERSION
+#endif
 #define _POSIX_THREAD_PROCESS_SHARED _POSIX_VERSION
 #define _POSIX_THREAD_SAFE_FUNCTIONS _POSIX_VERSION
 #define _POSIX_THREAD_ATTR_STACKADDR _POSIX_VERSION
diff --git a/libc-top-half/musl/src/env/__init_tls.c b/libc-top-half/musl/src/env/__init_tls.c
index ee785bc..ece8d24 100644
--- a/libc-top-half/musl/src/env/__init_tls.c
+++ b/libc-top-half/musl/src/env/__init_tls.c
@@ -16,6 +16,43 @@
 volatile int __thread_list_lock;
 
 #ifndef __wasilibc_unmodified_upstream
+
+/* These symbols are generated by wasm-ld. __stack_high/__stack_low
+ * symbols are only available in LLVM v16 and higher, therefore they're
+ * defined as weak symbols and if not available, __heap_base/__data_end
+ * is used instead.
+ *
+ * TODO: remove usage of __heap_base/__data_end for stack size calculation
+ * once we drop support for LLVM v15 and older.
+ */
+extern unsigned char __heap_base;
+extern unsigned char __data_end;
+extern unsigned char __global_base;
+extern weak unsigned char __stack_high;
+extern weak unsigned char __stack_low;
+
+static inline void setup_default_stack_size()
+{
+	ptrdiff_t stack_size;
+
+	if (&__stack_high)
+		stack_size = &__stack_high - &__stack_low;
+	else {
+		unsigned char *sp;
+		__asm__(
+			".globaltype __stack_pointer, i32\n"
+			"global.get __stack_pointer\n"
+			"local.set %0\n"
+			: "=r"(sp));
+		stack_size = sp > &__global_base ? &__heap_base - &__data_end : (ptrdiff_t)&__global_base;
+	}
+
+	if (stack_size > __default_stacksize)
+		__default_stacksize =
+			stack_size < DEFAULT_STACK_MAX ?
+			stack_size : DEFAULT_STACK_MAX;
+}
+
 void __wasi_init_tp() {
 	__init_tp((void *)__get_tp());
 }
@@ -31,6 +68,20 @@ int __init_tp(void *p)
 	if (!r) libc.can_do_threads = 1;
 	td->detach_state = DT_JOINABLE;
 	td->tid = __syscall(SYS_set_tid_address, &__thread_list_lock);
+#else
+	setup_default_stack_size();
+	td->detach_state = DT_JOINABLE;
+	/*
+	 * Initialize the TID to a value which doesn't conflict with
+	 * host-allocated TIDs, so that TID-based locks can work.
+	 *
+	 * Note:
+	 * - Host-allocated TIDs range from 1 to 0x1fffffff. (inclusive)
+	 * - __tl_lock and __lockfile uses TID 0 as "unlocked".
+	 * - __lockfile relies on the fact the most significant two bits
+	 *   of TIDs are 0.
+	 */
+	td->tid = 0x3fffffff;
 #endif
 	td->locale = &libc.global_locale;
 	td->robust_list.head = &td->robust_list.head;
diff --git a/libc-top-half/musl/src/include/pthread.h b/libc-top-half/musl/src/include/pthread.h
index 7167d3e..0aba813 100644
--- a/libc-top-half/musl/src/include/pthread.h
+++ b/libc-top-half/musl/src/include/pthread.h
@@ -7,7 +7,9 @@ hidden int __pthread_once(pthread_once_t *, void (*)(void));
 hidden void __pthread_testcancel(void);
 hidden int __pthread_setcancelstate(int, int *);
 hidden int __pthread_create(pthread_t *restrict, const pthread_attr_t *restrict, void *(*)(void *), void *restrict);
+#ifdef __wasilibc_unmodified_upstream
 hidden _Noreturn void __pthread_exit(void *);
+#endif
 hidden int __pthread_join(pthread_t, void **);
 hidden int __pthread_mutex_lock(pthread_mutex_t *);
 hidden int __pthread_mutex_trylock(pthread_mutex_t *);
diff --git a/libc-top-half/musl/src/internal/pthread_impl.h b/libc-top-half/musl/src/internal/pthread_impl.h
index a6d188b..1e7b974 100644
--- a/libc-top-half/musl/src/internal/pthread_impl.h
+++ b/libc-top-half/musl/src/internal/pthread_impl.h
@@ -216,7 +216,12 @@ extern hidden unsigned __default_stacksize;
 extern hidden unsigned __default_guardsize;
 
 #define DEFAULT_STACK_SIZE 131072
+#ifdef __wasilibc_unmodified_upstream
 #define DEFAULT_GUARD_SIZE 8192
+#else
+/* guard doesn't make much sense without mprotect. */
+#define DEFAULT_GUARD_SIZE 0
+#endif
 
 #define DEFAULT_STACK_MAX (8<<20)
 #define DEFAULT_GUARD_MAX (1<<20)
diff --git a/libc-top-half/musl/src/thread/__wait.c b/libc-top-half/musl/src/thread/__wait.c
index c0e4aac..7ffa987 100644
--- a/libc-top-half/musl/src/thread/__wait.c
+++ b/libc-top-half/musl/src/thread/__wait.c
@@ -48,7 +48,7 @@ void __wait(volatile int *addr, volatile int *waiters, int val, int priv)
 		__syscall(SYS_futex, addr, FUTEX_WAIT|priv, val, 0) != -ENOSYS
 		|| __syscall(SYS_futex, addr, FUTEX_WAIT, val, 0);
 #else
-		__wasilibc_futex_wait(addr, FUTEX_WAIT, val, 0);
+		__wasilibc_futex_wait(addr, FUTEX_WAIT, val, -1);
 #endif
 	}
 	if (waiters) a_dec(waiters);
diff --git a/libc-top-half/musl/src/thread/pthread_attr_get.c b/libc-top-half/musl/src/thread/pthread_attr_get.c
index f12ff44..0ac251c 100644
--- a/libc-top-half/musl/src/thread/pthread_attr_get.c
+++ b/libc-top-half/musl/src/thread/pthread_attr_get.c
@@ -17,6 +17,7 @@ int pthread_attr_getinheritsched(const pthread_attr_t *restrict a, int *restrict
 	return 0;
 }
 
+#ifdef __wasilibc_unmodified_upstream /* WASI has no CPU scheduling support. */
 int pthread_attr_getschedparam(const pthread_attr_t *restrict a, struct sched_param *restrict param)
 {
 	param->sched_priority = a->_a_prio;
@@ -28,6 +29,7 @@ int pthread_attr_getschedpolicy(const pthread_attr_t *restrict a, int *restrict
 	*policy = a->_a_policy;
 	return 0;
 }
+#endif
 
 int pthread_attr_getscope(const pthread_attr_t *restrict a, int *restrict scope)
 {
@@ -56,11 +58,13 @@ int pthread_barrierattr_getpshared(const pthread_barrierattr_t *restrict a, int
 	return 0;
 }
 
+#ifdef __wasilibc_unmodified_upstream /* Forward declaration of WASI's `__clockid` type. */
 int pthread_condattr_getclock(const pthread_condattr_t *restrict a, clockid_t *restrict clk)
 {
 	*clk = a->__attr & 0x7fffffff;
 	return 0;
 }
+#endif
 
 int pthread_condattr_getpshared(const pthread_condattr_t *restrict a, int *restrict pshared)
 {
diff --git a/libc-top-half/musl/src/thread/pthread_barrier_destroy.c b/libc-top-half/musl/src/thread/pthread_barrier_destroy.c
index 4ce0b2e..a347a2c 100644
--- a/libc-top-half/musl/src/thread/pthread_barrier_destroy.c
+++ b/libc-top-half/musl/src/thread/pthread_barrier_destroy.c
@@ -9,7 +9,9 @@ int pthread_barrier_destroy(pthread_barrier_t *b)
 			while ((v = b->_b_lock) & INT_MAX)
 				__wait(&b->_b_lock, 0, v, 0);
 		}
+#ifdef __wasilibc_unmodified_upstream /* WASI does not understand processes or locking between them. */
 		__vm_wait();
+#endif
 	}
 	return 0;
 }
diff --git a/libc-top-half/musl/src/thread/pthread_barrier_wait.c b/libc-top-half/musl/src/thread/pthread_barrier_wait.c
index cc2a8bb..0891b71 100644
--- a/libc-top-half/musl/src/thread/pthread_barrier_wait.c
+++ b/libc-top-half/musl/src/thread/pthread_barrier_wait.c
@@ -23,7 +23,9 @@ static int pshared_barrier_wait(pthread_barrier_t *b)
 			__wait(&b->_b_count, &b->_b_waiters2, v, 0);
 	}
 
+#ifdef __wasilibc_unmodified_upstream /* WASI does not understand processes or locking between them. */
 	__vm_lock();
+#endif
 
 	/* Ensure all threads have a vm lock before proceeding */
 	if (a_fetch_add(&b->_b_count, -1)==1-limit) {
@@ -44,7 +46,9 @@ static int pshared_barrier_wait(pthread_barrier_t *b)
 	if (v==INT_MIN+1 || (v==1 && w))
 		__wake(&b->_b_lock, 1, 0);
 
+#ifdef __wasilibc_unmodified_upstream /* WASI does not understand processes or locking between them. */
 	__vm_unlock();
+#endif
 
 	return ret;
 }
@@ -84,8 +88,12 @@ int pthread_barrier_wait(pthread_barrier_t *b)
 			a_spin();
 		a_inc(&inst->finished);
 		while (inst->finished == 1)
+#ifdef __wasilibc_unmodified_upstream
 			__syscall(SYS_futex,&inst->finished,FUTEX_WAIT|FUTEX_PRIVATE,1,0) != -ENOSYS
 			|| __syscall(SYS_futex,&inst->finished,FUTEX_WAIT,1,0);
+#else
+			__futexwait(&inst->finished, 1, 0);
+#endif
 		return PTHREAD_BARRIER_SERIAL_THREAD;
 	}
 
diff --git a/libc-top-half/musl/src/thread/pthread_create.c b/libc-top-half/musl/src/thread/pthread_create.c
index 1aa7be7..676e2cc 100644
--- a/libc-top-half/musl/src/thread/pthread_create.c
+++ b/libc-top-half/musl/src/thread/pthread_create.c
@@ -60,7 +60,11 @@ void __tl_sync(pthread_t td)
 	if (tl_lock_waiters) __wake(&__thread_list_lock, 1, 0);
 }
 
+#ifdef __wasilibc_unmodified_upstream
 _Noreturn void __pthread_exit(void *result)
+#else
+static void __pthread_exit(void *result)
+#endif
 {
 	pthread_t self = __pthread_self();
 	sigset_t set;
@@ -191,7 +195,7 @@ _Noreturn void __pthread_exit(void *result)
 		__tl_unlock();
 		free(self->map_base);
 		// Can't use `exit()` here, because it is too high level
-		for (;;) __wasi_proc_exit(0);
+		return;
 	}
 #endif
 
@@ -212,7 +216,6 @@ _Noreturn void __pthread_exit(void *result)
 	// do it manually here
 	__tl_unlock();
 	// Can't use `exit()` here, because it is too high level
-	for (;;) __wasi_proc_exit(0);
 #endif
 }
 
@@ -235,9 +238,14 @@ struct start_args {
 	volatile int control;
 	unsigned long sig_mask[_NSIG/8/sizeof(long)];
 #else
+	/*
+	 * Note: the offset of the "stack" and "tls_base" members
+	 * in this structure is hardcoded in wasi_thread_start.
+	 */
+	void *stack;
+	void *tls_base;
 	void *(*start_func)(void *);
 	void *start_arg;
-	void *tls_base;
 #endif
 };
 
@@ -271,32 +279,41 @@ static int start_c11(void *p)
 	return 0;
 }
 #else
-__attribute__((export_name("wasi_thread_start")))
-_Noreturn void wasi_thread_start(int tid, void *p)
+
+/*
+ * We want to ensure wasi_thread_start is linked whenever
+ * pthread_create is used. The following reference is to ensure that.
+ * Otherwise, the linker doesn't notice the dependency because
+ * wasi_thread_start is used indirectly via a wasm export.
+ */
+void wasi_thread_start(int tid, void *p);
+hidden void *__dummy_reference = wasi_thread_start;
+
+hidden void __wasi_thread_start_C(int tid, void *p)
 {
 	struct start_args *args = p;
-  	__asm__(".globaltype __tls_base, i32\n"
-			"local.get %0\n"
-			"global.set __tls_base\n"
-			:: "r"(args->tls_base));
 	pthread_t self = __pthread_self();
 	// Set the thread ID (TID) on the pthread structure. The TID is stored
 	// atomically since it is also stored by the parent thread; this way,
 	// whichever thread (parent or child) reaches this point first can proceed
 	// without waiting.
 	atomic_store((atomic_int *) &(self->tid), tid);
-	// Set the stack pointer.
-  	__asm__(".globaltype __stack_pointer, i32\n"
-			"local.get %0\n"
-			"global.set __stack_pointer\n"
-			:: "r"(self->stack));
 	// Execute the user's start function.
-	int (*start)(void*) = (int(*)(void*)) args->start_func;
-	__pthread_exit((void *)(uintptr_t)start(args->start_arg));
+	__pthread_exit(args->start_func(args->start_arg));
 }
 #endif
 
+#ifdef __wasilibc_unmodified_upstream
 #define ROUND(x) (((x)+PAGE_SIZE-1)&-PAGE_SIZE)
+#else
+/*
+ * As we allocate stack with malloc() instead of mmap/mprotect,
+ * there is no point to round it up to PAGE_SIZE.
+ * Instead, round up to a sane alignment.
+ * Note: PAGE_SIZE is rather big on WASM. (65536)
+ */
+#define ROUND(x) (((x)+16-1)&-16)
+#endif
 
 /* pthread_key_create.c overrides this */
 static volatile size_t dummy = 0;
@@ -484,6 +501,7 @@ int __pthread_create(pthread_t *restrict res, const pthread_attr_t *restrict att
 	/* Correct the stack size */
 	new->stack_size = stack - stack_limit;
 
+	args->stack = new->stack; /* just for convenience of asm trampoline */
 	args->start_func = entry;
 	args->start_arg = arg;
 	args->tls_base = (void*)new_tls_base;
@@ -561,5 +579,7 @@ fail:
 	return EAGAIN;
 }
 
+#ifdef __wasilibc_unmodified_upstream
 weak_alias(__pthread_exit, pthread_exit);
+#endif
 weak_alias(__pthread_create, pthread_create);
diff --git a/libc-top-half/musl/src/thread/wasm32/wasi_thread_start.s b/libc-top-half/musl/src/thread/wasm32/wasi_thread_start.s
new file mode 100644
index 0000000..0fe9854
--- /dev/null
+++ b/libc-top-half/musl/src/thread/wasm32/wasi_thread_start.s
@@ -0,0 +1,31 @@
+	.text
+
+	.export_name	wasi_thread_start, wasi_thread_start
+
+	.globaltype	__stack_pointer, i32
+	.globaltype	__tls_base, i32
+	.functype	__wasi_thread_start_C (i32, i32) -> ()
+
+	.hidden	wasi_thread_start
+	.globl	wasi_thread_start
+	.type	wasi_thread_start,@function
+
+wasi_thread_start:
+	.functype	wasi_thread_start (i32, i32) -> ()
+
+	# Set up the minimum C environment.
+	# Note: offsetof(start_arg, stack) == 0
+	local.get   1  # start_arg
+	i32.load    0  # stack
+	global.set  __stack_pointer
+
+	local.get   1  # start_arg
+	i32.load    4  # tls_base
+	global.set  __tls_base
+
+	# Make the C function do the rest of work.
+	local.get   0  # tid
+	local.get   1  # start_arg
+	call __wasi_thread_start_C
+
+	end_function