[wasm] add asyncify based setjmp, fiber, register scan emulation

configure.ac: setup build tools and register objects

main.c: wrap main with rb_wasm_rt_start to handle asyncify unwinds

tool/m4/ruby_wasm_tools.m4: setup default command based on WASI_SDK_PATH
environment variable. checks wasm-opt which is used for asyncify.

tool/wasm-clangw wasm/wasm-opt: a clang wrapper which replaces real
wasm-opt with do-nothing wasm-opt to avoid misoptimization before
asyncify. asyncify is performed at POSTLINK, but clang linker driver
tries to run optimization by wasm-opt unconditionally. inlining pass
at wasm level breaks asyncify's assumption, so should not optimize
before POSTLIK.

wasm/GNUmakefile.in: wasm specific rules to compile objects

configure.ac

@@ -42,6 +42,7 @@ m4_include([tool/m4/ruby_try_cflags.m4])dnl
 m4_include([tool/m4/ruby_try_cxxflags.m4])dnl
 m4_include([tool/m4/ruby_try_ldflags.m4])dnl
 m4_include([tool/m4/ruby_universal_arch.m4])dnl
+m4_include([tool/m4/ruby_wasm_tools.m4])dnl
 m4_include([tool/m4/ruby_werror_flag.m4])dnl
 
 AC_ARG_VAR([cflags], [additional CFLAGS (ignored when CFLAGS is given)])dnl
@@ -136,6 +137,9 @@ AS_IF([test ! -z "$ac_cv_prog_CC" -a ! -z "$CC" -a "$CC" != "$ac_cv_prog_CC"], [
   AC_MSG_ERROR(cached CC is different -- throw away $cache_file
 (it is also a good idea to do 'make clean' before compiling))
 ])
+
+RUBY_WASM_TOOLS
+
 AS_CASE(["${build_os}"],
 [linux*|cygwin*|msys*], [
     # Naruse prefers GCC on Linux
@@ -402,6 +406,15 @@ AS_CASE(["$build_os"],
     ])
     rm -fr conftest*
   ])
+AS_CASE(["$target_os"],
+  [wasi*], [
+    # Clang linker automatically uses wasm-opt with -O if it found.
+    # https://github.com/llvm/llvm-project/blob/812828984c10857a4cd260eb638c52a4411f9143/clang/lib/Driver/ToolChains/WebAssembly.cpp#L95-L118
+    # However optimization before asyncify causes misoptimization,
+    # so wrap clang to insert our fake wasm-opt, which does nothing, in PATH.
+    CC_WRAPPER=`cd -P "${tooldir}" && pwd`/wasm-clangw
+    CC="$CC_WRAPPER $CC"
+  ])
 
 cc_version=
 for option in --version -v -V -qversion; do
@@ -1218,6 +1231,7 @@ main()
 [wasi*],[	LIBS="-lm -lwasi-emulated-mman -lwasi-emulated-signal -lwasi-emulated-getpid -lwasi-emulated-process-clocks $LIBS"
 		RUBY_APPEND_OPTIONS(CFLAGS, -D_WASI_EMULATED_SIGNAL -D_WASI_EMULATED_MMAN -D_WASI_EMULATED_GETPID -D_WASI_EMULATED_PROCESS_CLOCKS)
 		RUBY_APPEND_OPTIONS(CPPFLAGS, -D_WASI_EMULATED_SIGNAL -D_WASI_EMULATED_MMAN -D_WASI_EMULATED_GETPID -D_WASI_EMULATED_PROCESS_CLOCKS)
+		POSTLINK="\$(WASMOPT) --asyncify \$(wasmoptflags) --pass-arg=asyncify-ignore-imports -o \$@ \$@${POSTLINK:+; $POSTLINK}"
 		]
 [	LIBS="-lm $LIBS"])
 : ${ORIG_LIBS=$LIBS}
@@ -3790,6 +3804,16 @@ AS_CASE(["$target_os"],
 	    LIBRUBY='lib$(RUBY_SO_NAME).a'
 	    LIBRUBYARG='-l$(RUBY_SO_NAME)'
 	])
+	],
+    [wasi*], [
+        FIRSTMAKEFILE=GNUmakefile:wasm/GNUmakefile.in
+        AC_LIBOBJ([wasm/runtime])
+        AC_LIBOBJ([wasm/fiber])
+        AC_LIBOBJ([wasm/machine])
+        AC_LIBOBJ([wasm/setjmp])
+        AC_LIBOBJ([wasm/machine_core])
+        AC_LIBOBJ([wasm/setjmp_core])
+        PLATFORM_DIR=wasm
 ])
 
 MINIOBJS="$MINIDLNOBJ"

main.c

@@ -31,7 +31,7 @@
 #endif
 
 int
-main(int argc, char **argv)
+rb_main(int argc, char **argv)
 {
 #ifdef RUBY_DEBUG_ENV
     ruby_set_debug_option(getenv("RUBY_DEBUG"));
@@ -47,3 +47,12 @@ main(int argc, char **argv)
 	return ruby_run_node(ruby_options(argc, argv));
     }
 }
+
+int main(int argc, char **argv) {
+#if defined(__wasm__) && !defined(__EMSCRIPTEN__)
+  int rb_wasm_rt_start(int (main)(int argc, char **argv), int argc, char **argv);
+  return rb_wasm_rt_start(rb_main, argc, argv);
+#else
+  return rb_main(argc, argv);
+#endif
+}

tool/m4/ruby_wasm_tools.m4

@@ -0,0 +1,21 @@
+dnl -*- Autoconf -*-
+AC_DEFUN([RUBY_WASM_TOOLS],
+[AS_CASE(["$target_os"],
+[wasi*], [
+    AC_CHECK_TOOL(WASMOPT, wasm-opt)
+    AS_IF([test x"${WASMOPT}" = x], [
+        AC_MSG_ERROR([wasm-opt is required])
+    ])
+    AC_SUBST(wasmoptflags)
+    : ${wasmoptflags=-O3}
+
+    AC_MSG_CHECKING([wheather \$WASI_SDK_PATH is set])
+    AS_IF([test x"${WASI_SDK_PATH}" = x], [AC_MSG_RESULT([no])], [
+        AC_MSG_RESULT([yes])
+        CC="${WASI_SDK_PATH}/bin/clang"
+        LD="${WASI_SDK_PATH}/bin/clang"
+        AR="${WASI_SDK_PATH}/bin/llvm-ar"
+        RANLIB="${WASI_SDK_PATH}/bin/llvm-ranlib"
+    ])
+])
+])dnl

tool/wasm-clangw

@@ -0,0 +1,9 @@
+#!/bin/sh
+# A Clang wrapper script to fake the clang linker driver.
+# Clang linker automatically uses wasm-opt with -O if it found.
+# However optimization before asyncify causes misoptimization,
+# so wrap clang to insert our fake wasm-opt, which does nothing, in PATH.
+
+src_dir="$(cd "$(dirname "$0")/../wasm" && pwd)"
+export PATH="$src_dir:$PATH"
+exec "$@"

wasm/GNUmakefile.in

@@ -0,0 +1,19 @@
+include Makefile
+include $(srcdir)/template/GNUmakefile.in
+
+wasmdir = $(srcdir)/wasm
+GNUmakefile: $(wasmdir)/GNUmakefile.in
+WASMOPT = @WASMOPT@
+wasmoptflags = @wasmoptflags@
+
+WASM_OBJS  = $(wasmdir)/machine_core.o $(wasmdir)/machine.o $(wasmdir)/setjmp.o $(wasmdir)/setjmp_core.o $(wasmdir)/fiber.o $(wasmdir)/runtime.o
+
+wasm/fiber.$(OBJEXT): $(wasmdir)/fiber.c $(wasmdir)/fiber.h $(wasmdir)/asyncify.h $(PLATFORM_D)
+wasm/machine.$(OBJEXT): $(wasmdir)/machine.c $(srcdir)/wasm/machine.h $(wasmdir)/asyncify.h $(PLATFORM_D)
+wasm/setjmp.$(OBJEXT): $(wasmdir)/setjmp.c $(wasmdir)/setjmp.h $(wasmdir)/machine.h $(wasmdir)/asyncify.h $(PLATFORM_D)
+wasm/runtime.$(OBJEXT): $(wasmdir)/runtime.c $(wasmdir)/machine.h $(wasmdir)/asyncify.h $(wasmdir)/setjmp.h $(PLATFORM_D)
+
+wasm/%.$(OBJEXT): $(wasmdir)/%.S $(PLATFORM_D)
+	@$(ECHO) compiling $<
+	$(Q) $(CC) $(CFLAGS) $(COUTFLAG)$@ -c $<
+

wasm/asyncify.h

@@ -0,0 +1,13 @@
+#ifndef RB_WASM_SUPPORT_ASYNCIFY_H
+#define RB_WASM_SUPPORT_ASYNCIFY_H
+
+__attribute__((import_module("asyncify"), import_name("start_unwind")))
+void asyncify_start_unwind(void *buf);
+__attribute__((import_module("asyncify"), import_name("stop_unwind")))
+void asyncify_stop_unwind(void);
+__attribute__((import_module("asyncify"), import_name("start_rewind")))
+void asyncify_start_rewind(void *buf);
+__attribute__((import_module("asyncify"), import_name("stop_rewind")))
+void asyncify_stop_rewind(void);
+
+#endif

wasm/fiber.c

@@ -0,0 +1,83 @@
+/*
+ This is a ucontext-like userland context switching API for WebAssembly based on Binaryen's Asyncify.
+
+ * NOTE:
+ * This mechanism doesn't take care of stack state. Just save and restore program counter and
+ * registers (rephrased as locals by Wasm term). So use-site need to save and restore the C stack pointer.
+ * This Asyncify based implementation is not much efficient and will be replaced with future stack-switching feature.
+ */
+
+#include <stdlib.h>
+#include "wasm/fiber.h"
+#include "wasm/asyncify.h"
+
+#ifdef RB_WASM_ENABLE_DEBUG_LOG
+# include <stdio.h>
+# define RB_WASM_DEBUG_LOG(...) fprintf(stderr, __VA_ARGS__)
+#else
+# define RB_WASM_DEBUG_LOG(...)
+#endif
+
+void
+rb_wasm_init_context(rb_wasm_fiber_context *fcp, void (*func)(void *, void *), void *arg0, void *arg1)
+{
+    fcp->asyncify_buf.top = &fcp->asyncify_buf.buffer[0];
+    fcp->asyncify_buf.end = &fcp->asyncify_buf.buffer[WASM_FIBER_STACK_BUFFER_SIZE];
+    fcp->is_rewinding = false;
+    fcp->is_started = false;
+    fcp->entry_point = func;
+    fcp->arg0 = arg0;
+    fcp->arg1 = arg1;
+    RB_WASM_DEBUG_LOG("[%s] fcp->asyncify_buf %p\n", __func__, &fcp->asyncify_buf);
+}
+
+static rb_wasm_fiber_context *_rb_wasm_active_next_fiber;
+
+void
+rb_wasm_swapcontext(rb_wasm_fiber_context *ofcp, rb_wasm_fiber_context *fcp)
+{
+    RB_WASM_DEBUG_LOG("[%s] enter ofcp = %p fcp = %p\n", __func__, ofcp, fcp);
+    if (ofcp->is_rewinding) {
+        asyncify_stop_rewind();
+        ofcp->is_rewinding = false;
+        return;
+    }
+    _rb_wasm_active_next_fiber = fcp;
+    RB_WASM_DEBUG_LOG("[%s] start unwinding asyncify_buf = %p\n", __func__, &ofcp->asyncify_buf);
+    asyncify_start_unwind(&ofcp->asyncify_buf);
+}
+
+void *
+rb_wasm_handle_fiber_unwind(void (**new_fiber_entry)(void *, void *),
+                            void **arg0, void **arg1, bool *is_new_fiber_started)
+{
+    rb_wasm_fiber_context *next_fiber;
+    if (!_rb_wasm_active_next_fiber) {
+        RB_WASM_DEBUG_LOG("[%s] no next fiber\n", __func__);
+        *is_new_fiber_started = false;
+        return NULL;
+    }
+
+    next_fiber = _rb_wasm_active_next_fiber;
+    _rb_wasm_active_next_fiber = NULL;
+
+    RB_WASM_DEBUG_LOG("[%s] next_fiber->asyncify_buf = %p\n", __func__, &next_fiber->asyncify_buf);
+
+    *new_fiber_entry = next_fiber->entry_point;
+    *arg0 = next_fiber->arg0;
+    *arg1 = next_fiber->arg1;
+
+    if (!next_fiber->is_started) {
+        RB_WASM_DEBUG_LOG("[%s] new fiber started\n", __func__);
+        // start a new fiber if not started yet.
+        next_fiber->is_started = true;
+        *is_new_fiber_started = true;
+        return NULL;
+    } else {
+        RB_WASM_DEBUG_LOG("[%s] resume a fiber\n", __func__);
+        // resume a fiber again
+        next_fiber->is_rewinding = true;
+        *is_new_fiber_started = false;
+        return &next_fiber->asyncify_buf;
+    }
+}

wasm/fiber.h

@@ -0,0 +1,43 @@
+#ifndef RB_WASM_SUPPORT_FIBER_H
+#define RB_WASM_SUPPORT_FIBER_H
+
+#include <stdbool.h>
+
+#ifndef WASM_FIBER_STACK_BUFFER_SIZE
+# define WASM_FIBER_STACK_BUFFER_SIZE 6144
+#endif
+
+struct __rb_wasm_asyncify_fiber_ctx {
+    void* top;
+    void* end;
+    char buffer[WASM_FIBER_STACK_BUFFER_SIZE];
+};
+
+// Fiber execution context needed to perform context switch
+typedef struct {
+    // Fiber entry point called when the fiber started for the first time.
+    // NULL if the entry point is main
+    void (*entry_point)(void *, void *);
+    // Opaque argument pointers passed to the entry point function
+    void *arg0, *arg1;
+
+    // Internal asyncify buffer space
+    struct __rb_wasm_asyncify_fiber_ctx asyncify_buf;
+
+    bool is_rewinding;
+    bool is_started;
+} rb_wasm_fiber_context;
+
+// Initialize a given fiber context to be ready to pass to `rb_wasm_swapcontext`
+void rb_wasm_init_context(rb_wasm_fiber_context *fcp, void (*func)(void *, void *), void *arg0, void *arg1);
+
+// Swap the execution control with `target_fiber` and save the current context in `old_fiber`
+// NOTE: `old_fiber` must be the current executing fiber context
+void rb_wasm_swapcontext(rb_wasm_fiber_context *old_fiber, rb_wasm_fiber_context *target_fiber);
+
+// Returns the Asyncify buffer of next fiber if unwound for fiber context switch.
+// Used by the top level Asyncify handling in wasm/runtime.c
+void *rb_wasm_handle_fiber_unwind(void (**new_fiber_entry)(void *, void *),
+                                  void **arg0, void **arg1, bool *is_new_fiber_started);
+
+#endif

wasm/machine.c

@@ -0,0 +1,62 @@
+#include <stdlib.h>
+#include "wasm/machine.h"
+#include "wasm/asyncify.h"
+
+#ifndef WASM_SCAN_STACK_BUFFER_SIZE
+# define WASM_SCAN_STACK_BUFFER_SIZE 6144
+#endif
+
+struct asyncify_buf {
+    void *top;
+    void *end;
+    uint8_t buffer[WASM_SCAN_STACK_BUFFER_SIZE];
+};
+
+static void
+init_asyncify_buf(struct asyncify_buf* buf)
+{
+    buf->top = &buf->buffer[0];
+    buf->end = &buf->buffer[WASM_SCAN_STACK_BUFFER_SIZE];
+}
+
+static void *_rb_wasm_active_scan_buf = NULL;
+
+void
+rb_wasm_scan_locals(rb_wasm_scan_func scan)
+{
+    static struct asyncify_buf buf;
+    static int spilling = 0;
+    if (!spilling) {
+        spilling = 1;
+        init_asyncify_buf(&buf);
+        _rb_wasm_active_scan_buf = &buf;
+        asyncify_start_unwind(&buf);
+    } else {
+        asyncify_stop_rewind();
+        spilling = 0;
+        _rb_wasm_active_scan_buf = NULL;
+        scan(buf.top, buf.end);
+    }
+}
+
+static void *rb_wasm_stack_base = NULL;
+
+__attribute__((constructor))
+int
+rb_wasm_record_stack_base(void)
+{
+    rb_wasm_stack_base = rb_wasm_get_stack_pointer();
+    return 0;
+}
+
+void
+_rb_wasm_scan_stack(rb_wasm_scan_func scan, void *current)
+{
+    scan(current, rb_wasm_stack_base);
+}
+
+void *
+rb_wasm_handle_scan_unwind(void)
+{
+    return _rb_wasm_active_scan_buf;
+}

wasm/machine.h

@@ -0,0 +1,26 @@
+#ifndef RB_WASM_SUPPORT_MACHINE_H
+#define RB_WASM_SUPPORT_MACHINE_H
+
+// Function pointer used as scan callbacks
+typedef void (*rb_wasm_scan_func)(void*, void*);
+
+// Scan WebAssembly locals in the all call stack (like registers) spilled by Asyncify
+// Used by conservative GC
+void rb_wasm_scan_locals(rb_wasm_scan_func scan);
+
+// Scan userland C-stack memory space in WebAssembly. Used by conservative GC
+#define rb_wasm_scan_stack(scan) _rb_wasm_scan_stack((scan), rb_wasm_get_stack_pointer())
+void _rb_wasm_scan_stack(rb_wasm_scan_func scan, void *current);
+
+
+// Get the current stack pointer
+void *rb_wasm_get_stack_pointer(void);
+
+// Set the current stack pointer
+void rb_wasm_set_stack_pointer(void *sp);
+
+// Returns the Asyncify buffer of next rewinding if unwound for spilling locals.
+// Used by the top level Asyncify handling in wasm/runtime.c
+void *rb_wasm_handle_scan_unwind(void);
+
+#endif

wasm/machine_core.S

@@ -0,0 +1,25 @@
+	# extern int __stack_pointer;
+	.globaltype __stack_pointer, i32
+
+	# NOTE: Implement this in raw assembly to avoid stack pointer
+	#       operations in C-prologue and epilogue.
+
+	# void *rb_wasm_get_stack_pointer(void);
+	.section	.text.rb_wasm_get_stack_pointer,"",@
+	.globl	rb_wasm_get_stack_pointer
+	.type	rb_wasm_get_stack_pointer,@function
+rb_wasm_get_stack_pointer:
+	.functype	rb_wasm_get_stack_pointer () -> (i32)
+	global.get	__stack_pointer
+	end_function
+
+	# void rb_wasm_set_stack_pointer(void *sp);
+	.section	.text.rb_wasm_set_stack_pointer,"",@
+	.globl	rb_wasm_set_stack_pointer
+	.type	rb_wasm_set_stack_pointer,@function
+rb_wasm_set_stack_pointer:
+	.functype	rb_wasm_set_stack_pointer (i32) -> ()
+	local.get	0
+	global.set	__stack_pointer
+	end_function
+

wasm/runtime.c

@@ -0,0 +1,47 @@
+#include "wasm/machine.h"
+#include "wasm/setjmp.h"
+#include "wasm/fiber.h"
+#include "wasm/asyncify.h"
+#include <stdlib.h>
+
+int rb_wasm_rt_start(int (main)(int argc, char **argv), int argc, char **argv) {
+  int result;
+  void *asyncify_buf;
+
+  bool new_fiber_started = false;
+  void *arg0 = NULL, *arg1 = NULL;
+  void (*fiber_entry_point)(void *, void *) = NULL;
+
+  while (1) {
+    if (fiber_entry_point) {
+      fiber_entry_point(arg0, arg1);
+    } else {
+      result = main(argc, argv);
+    }
+
+    // NOTE: it's important to call 'asyncify_stop_unwind' here instead in rb_wasm_handle_jmp_unwind
+    // because unless that, Asyncify inserts another unwind check here and it unwinds to the root frame.
+    asyncify_stop_unwind();
+
+    if ((asyncify_buf = rb_wasm_handle_jmp_unwind()) != NULL) {
+      asyncify_start_rewind(asyncify_buf);
+      continue;
+    }
+    if ((asyncify_buf = rb_wasm_handle_scan_unwind()) != NULL) {
+      asyncify_start_rewind(asyncify_buf);
+      continue;
+    }
+
+    asyncify_buf = rb_wasm_handle_fiber_unwind(&fiber_entry_point, &arg0, &arg1, &new_fiber_started);
+    // Newly starting fiber doesn't have asyncify buffer yet, so don't rewind it for the first time entry
+    if (asyncify_buf) {
+      asyncify_start_rewind(asyncify_buf);
+      continue;
+    } else if (new_fiber_started) {
+      continue;
+    }
+
+    break;
+  }
+  return result;
+}

wasm/setjmp.c

@@ -0,0 +1,134 @@
+/*
+ This is a WebAssembly userland setjmp/longjmp implementation based on Binaryen's Asyncify.
+ Inspired by Alon Zakai's snippet released under the MIT License:
+ * https://github.com/kripken/talks/blob/991fb1e4b6d7e4b0ea6b3e462d5643f11d422771/jmp.c
+
+ WebAssembly doesn't have context-switching mechanism for now, so emulate it by Asyncify,
+ which transforms WebAssembly binary to unwind/rewind the execution point and store/restore
+ locals.
+
+ The basic concept of this implementation is:
+ 1. setjmp captures the current execution context by unwinding to the root frame, then immediately
+    rewind to the setjmp call using the captured context. The context is saved in jmp_buf.
+ 2. longjmp unwinds to the root frame and rewinds to a setjmp call re-using a passed jmp_buf.
+
+ This implementation also supports switching context across different call stack (non-standard)
+
+ This approach is good at behavior reproducibility and self-containedness compared to Emscripten's
+ JS exception approach. However this is super expensive because Asyncify inserts many glue code to
+ control execution point in userland.
+
+ This implementation will be replaced with future stack-switching feature.
+ */
+#include <stdint.h>
+#include <stdlib.h>
+#include <assert.h>
+#include <stdbool.h>
+#include "wasm/asyncify.h"
+#include "wasm/machine.h"
+#include "wasm/setjmp.h"
+
+#ifdef RB_WASM_ENABLE_DEBUG_LOG
+# include <stdio.h>
+# define RB_WASM_DEBUG_LOG(...) fprintf(stderr, __VA_ARGS__)
+#else
+# define RB_WASM_DEBUG_LOG(...)
+#endif
+
+enum rb_wasm_jmp_buf_state {
+    // Initial state
+    JMP_BUF_STATE_INITIALIZED = 0,
+    // Unwinding to the root or rewinding to the setjmp call
+    // to capture the current execution context
+    JMP_BUF_STATE_CAPTURING   = 1,
+    // Ready for longjmp
+    JMP_BUF_STATE_CAPTURED    = 2,
+    // Unwinding to the root or rewinding to the setjmp call
+    // to restore the execution context
+    JMP_BUF_STATE_RETURNING   = 3,
+};
+
+void
+async_buf_init(struct __rb_wasm_asyncify_jmp_buf* buf)
+{
+    buf->top = &buf->buffer[0];
+    buf->end = &buf->buffer[WASM_SETJMP_STACK_BUFFER_SIZE];
+}
+
+// Global unwinding/rewinding jmpbuf state
+static rb_wasm_jmp_buf *_rb_wasm_active_jmpbuf;
+
+__attribute__((noinline))
+int
+_rb_wasm_setjmp_internal(rb_wasm_jmp_buf *env)
+{
+    RB_WASM_DEBUG_LOG("[%s] env = %p, env->state = %d, _rb_wasm_active_jmpbuf = %p\n", __func__, env, env->state, _rb_wasm_active_jmpbuf);
+    switch (env->state) {
+    case JMP_BUF_STATE_INITIALIZED: {
+        RB_WASM_DEBUG_LOG("[%s] JMP_BUF_STATE_INITIALIZED\n", __func__);
+        env->state = JMP_BUF_STATE_CAPTURING;
+        env->payload = 0;
+        _rb_wasm_active_jmpbuf = env;
+        async_buf_init(&env->setjmp_buf);
+        asyncify_start_unwind(&env->setjmp_buf);
+        return -1; // return a dummy value
+    }
+    case JMP_BUF_STATE_CAPTURING: {
+        asyncify_stop_rewind();
+        RB_WASM_DEBUG_LOG("[%s] JMP_BUF_STATE_CAPTURING\n", __func__);
+        env->state = JMP_BUF_STATE_CAPTURED;
+        _rb_wasm_active_jmpbuf = NULL;
+        return 0;
+    }
+    case JMP_BUF_STATE_RETURNING: {
+        asyncify_stop_rewind();
+        RB_WASM_DEBUG_LOG("[%s] JMP_BUF_STATE_RETURNING\n", __func__);
+        env->state = JMP_BUF_STATE_CAPTURED;
+        _rb_wasm_active_jmpbuf = NULL;
+        return env->payload;
+    }
+    default:
+        assert(0 && "unexpected state");
+    }
+    return 0;
+}
+
+void
+_rb_wasm_longjmp(rb_wasm_jmp_buf* env, int value)
+{
+    RB_WASM_DEBUG_LOG("[%s] env = %p, env->state = %d, value = %d\n", __func__, env, env->state, value);
+    assert(env->state == JMP_BUF_STATE_CAPTURED);
+    assert(value != 0);
+    env->state = JMP_BUF_STATE_RETURNING;
+    env->payload = value;
+    _rb_wasm_active_jmpbuf = env;
+    async_buf_init(&env->longjmp_buf);
+    asyncify_start_unwind(&env->longjmp_buf);
+}
+
+void *
+rb_wasm_handle_jmp_unwind(void)
+{
+    RB_WASM_DEBUG_LOG("[%s] _rb_wasm_active_jmpbuf = %p\n", __func__, _rb_wasm_active_jmpbuf);
+    if (!_rb_wasm_active_jmpbuf) {
+	return NULL;
+    }
+
+    switch (_rb_wasm_active_jmpbuf->state) {
+    case JMP_BUF_STATE_CAPTURING: {
+        RB_WASM_DEBUG_LOG("[%s] JMP_BUF_STATE_CAPTURING\n", __func__);
+        // save the captured Asyncify stack top
+        _rb_wasm_active_jmpbuf->dst_buf_top = _rb_wasm_active_jmpbuf->setjmp_buf.top;
+        break;
+    }
+    case JMP_BUF_STATE_RETURNING: {
+        RB_WASM_DEBUG_LOG("[%s] JMP_BUF_STATE_RETURNING\n", __func__);
+        // restore the saved Asyncify stack top
+        _rb_wasm_active_jmpbuf->setjmp_buf.top = _rb_wasm_active_jmpbuf->dst_buf_top;
+        break;
+    }
+    default:
+        assert(0 && "unexpected state");
+    }
+    return &_rb_wasm_active_jmpbuf->setjmp_buf;
+}

wasm/setjmp.h

@@ -0,0 +1,61 @@
+#ifndef RB_WASM_SUPPORT_SETJMP_H
+#define RB_WASM_SUPPORT_SETJMP_H
+
+#include "ruby/internal/config.h"
+#include <stdbool.h>
+
+#ifndef WASM_SETJMP_STACK_BUFFER_SIZE
+# define WASM_SETJMP_STACK_BUFFER_SIZE 6144
+#endif
+
+struct __rb_wasm_asyncify_jmp_buf {
+    void* top;
+    void* end;
+    char buffer[WASM_SETJMP_STACK_BUFFER_SIZE];
+};
+
+typedef struct {
+    // Internal Asyncify buffer space to save execution context
+    struct __rb_wasm_asyncify_jmp_buf setjmp_buf;
+    // Internal Asyncify buffer space used while unwinding from longjmp
+    // but never used for rewinding.
+    struct __rb_wasm_asyncify_jmp_buf longjmp_buf;
+    // Used to save top address of Asyncify stack `setjmp_buf`, which is
+    // overwritten during first rewind.
+    void *dst_buf_top;
+    // A payload value given by longjmp and returned by setjmp for the second time
+    int payload;
+    // Internal state field
+    int state;
+} rb_wasm_jmp_buf;
+
+// noinline to avoid breaking Asyncify assumption
+NOINLINE(int _rb_wasm_setjmp(rb_wasm_jmp_buf *env));
+NOINLINE(void _rb_wasm_longjmp(rb_wasm_jmp_buf *env, int payload));
+
+#define rb_wasm_setjmp(env) ((env).state = 0, _rb_wasm_setjmp(&(env)))
+
+// NOTE: Why is `_rb_wasm_longjmp` not `noreturn`? Why put `unreachable` in the call site?
+// Asyncify expects that `_rb_wasm_longjmp` returns its control, and Asyncify inserts a return
+// for unwinding after the call. This means that "`_rb_wasm_longjmp` returns its control but the
+// next line in the caller (C level) won't be executed".
+// On the other hand, `noreturn` means the callee won't return its control to the caller,
+// so compiler can assume that a function with the attribute won't reach the end of the function.
+// Therefore `_rb_wasm_longjmp`'s semantics is not exactly same as `noreturn`.
+#define rb_wasm_longjmp(env, payload) (_rb_wasm_longjmp(&env, payload), __builtin_unreachable())
+
+// Returns the Asyncify buffer of next rewinding if unwound for setjmp capturing or longjmp.
+// Used by the top level Asyncify handling in wasm/runtime.c
+void *rb_wasm_handle_jmp_unwind(void);
+
+
+//
+// POSIX-compatible declarations
+//
+
+typedef rb_wasm_jmp_buf jmp_buf;
+
+#define setjmp(env) rb_wasm_setjmp(env)
+#define longjmp(env, payload) rb_wasm_longjmp(env, payload)
+
+#endif

wasm/setjmp_core.S

@@ -0,0 +1,27 @@
+	# extern int _rb_wasm_setjmp_internal(rb_wasm_jmp_buf *env);
+	.functype	_rb_wasm_setjmp_internal (i32) -> (i32)
+	# extern int __stack_pointer;
+	.globaltype 	__stack_pointer, i32
+
+	# A wrapper of _rb_wasm_setjmp_internal to save and restore stack pointer
+	# This cannot be implemented in C because there is no way to manipulate stack pointer
+	# without C-epilogue.
+
+	# extern int _rb_wasm_setjmp(rb_wasm_jmp_buf *env);
+	.section	.text._rb_wasm_setjmp,"",@
+	.globl		_rb_wasm_setjmp
+	.type		_rb_wasm_setjmp,@function
+_rb_wasm_setjmp:
+	.functype	_rb_wasm_setjmp (i32) -> (i32)
+	.local		i32, i32
+	# save sp (this local is stored in asyncify stack and restored when rewinding)
+	global.get	__stack_pointer
+	local.set	1
+
+	local.get	0
+	call		_rb_wasm_setjmp_internal
+
+	# restore sp
+	local.get	1
+	global.set	__stack_pointer
+	end_function

wasm/wasm-opt

@@ -0,0 +1,36 @@
+#!/bin/sh
+# A fake wasm-opt, which does nothing at all
+# See also: tool/wasm-clangw
+
+set -e
+input=
+output=
+while [ $# -ne 0 ]; do
+  case "$1" in
+    -o)
+      shift
+      output=$1
+  ;;
+    -*)
+      # ignore other options
+  ;;
+    *)
+      input=$1
+  ;;
+  esac
+  shift
+done
+
+if [ -z "$input" ]; then
+  echo "missing input binary"
+  exit 1
+fi
+
+if [ -z "$output" ]; then
+  echo "missing output binary"
+  exit 1
+fi
+
+if [ "$input" != "$output" ]; then
+  cp "$input" "$output"
+fi