Merge pull request #1255 from odin-lang/wasi-wasm

`wasi_wasm32` support

core/fmt/fmt.odin

@@ -2,7 +2,6 @@ package fmt
 
 import "core:math/bits"
 import "core:mem"
-import "core:os"
 import "core:io"
 import "core:reflect"
 import "core:runtime"
@@ -62,38 +61,6 @@ register_user_formatter :: proc(id: typeid, formatter: User_Formatter) -> Regist
 }
 
 
-fprint :: proc(fd: os.Handle, args: ..any, sep := " ") -> int {
-	w := io.to_writer(os.stream_from_handle(fd))
-	return wprint(w=w, args=args, sep=sep)
-}
-
-fprintln :: proc(fd: os.Handle, args: ..any, sep := " ") -> int {
-	w := io.to_writer(os.stream_from_handle(fd))
-	return wprintln(w=w, args=args, sep=sep)
-}
-fprintf :: proc(fd: os.Handle, fmt: string, args: ..any) -> int {
-	w := io.to_writer(os.stream_from_handle(fd))
-	return wprintf(w, fmt, ..args)
-}
-fprint_type :: proc(fd: os.Handle, info: ^runtime.Type_Info) -> (n: int, err: io.Error) {
-	w := io.to_writer(os.stream_from_handle(fd))
-	return wprint_type(w, info)
-}
-fprint_typeid :: proc(fd: os.Handle, id: typeid) -> (n: int, err: io.Error) {
-	w := io.to_writer(os.stream_from_handle(fd))
-	return wprint_typeid(w, id)
-}
-
-// print* procedures return the number of bytes written
-print   :: proc(args: ..any, sep := " ") -> int { return fprint(fd=os.stdout, args=args, sep=sep) }
-println :: proc(args: ..any, sep := " ") -> int { return fprintln(fd=os.stdout, args=args, sep=sep) }
-printf  :: proc(fmt: string, args: ..any) -> int { return fprintf(os.stdout, fmt, ..args) }
-
-eprint   :: proc(args: ..any, sep := " ") -> int { return fprint(fd=os.stderr, args=args, sep=sep) }
-eprintln :: proc(args: ..any, sep := " ") -> int { return fprintln(fd=os.stderr, args=args, sep=sep) }
-eprintf  :: proc(fmt: string, args: ..any) -> int { return fprintf(os.stderr, fmt, ..args) }
-
-
 // aprint* procedures return a string that was allocated with the current context
 // They must be freed accordingly
 aprint :: proc(args: ..any, sep := " ") -> string {

core/fmt/fmt_os.odin

@@ -0,0 +1,37 @@
+//+build !freestanding
+package fmt
+
+import "core:runtime"
+import "core:os"
+import "core:io"
+
+fprint :: proc(fd: os.Handle, args: ..any, sep := " ") -> int {
+	w := io.to_writer(os.stream_from_handle(fd))
+	return wprint(w=w, args=args, sep=sep)
+}
+
+fprintln :: proc(fd: os.Handle, args: ..any, sep := " ") -> int {
+	w := io.to_writer(os.stream_from_handle(fd))
+	return wprintln(w=w, args=args, sep=sep)
+}
+fprintf :: proc(fd: os.Handle, fmt: string, args: ..any) -> int {
+	w := io.to_writer(os.stream_from_handle(fd))
+	return wprintf(w, fmt, ..args)
+}
+fprint_type :: proc(fd: os.Handle, info: ^runtime.Type_Info) -> (n: int, err: io.Error) {
+	w := io.to_writer(os.stream_from_handle(fd))
+	return wprint_type(w, info)
+}
+fprint_typeid :: proc(fd: os.Handle, id: typeid) -> (n: int, err: io.Error) {
+	w := io.to_writer(os.stream_from_handle(fd))
+	return wprint_typeid(w, id)
+}
+
+// print* procedures return the number of bytes written
+print   :: proc(args: ..any, sep := " ") -> int { return fprint(fd=os.stdout, args=args, sep=sep) }
+println :: proc(args: ..any, sep := " ") -> int { return fprintln(fd=os.stdout, args=args, sep=sep) }
+printf  :: proc(fmt: string, args: ..any) -> int { return fprintf(os.stdout, fmt, ..args) }
+
+eprint   :: proc(args: ..any, sep := " ") -> int { return fprint(fd=os.stderr, args=args, sep=sep) }
+eprintln :: proc(args: ..any, sep := " ") -> int { return fprintln(fd=os.stderr, args=args, sep=sep) }
+eprintf  :: proc(fmt: string, args: ..any) -> int { return fprintf(os.stderr, fmt, ..args) }

core/os/os_js_wasm32.odin

@@ -1,70 +0,0 @@
-package os
-
-Handle :: distinct i32;
-Errno :: distinct i32;
-
-ERROR_NONE :: Errno(0);
-
-O_RDONLY   :: 0x00000;
-O_WRONLY   :: 0x00001;
-O_RDWR     :: 0x00002;
-O_CREATE   :: 0x00040;
-O_EXCL     :: 0x00080;
-O_NOCTTY   :: 0x00100;
-O_TRUNC    :: 0x00200;
-O_NONBLOCK :: 0x00800;
-O_APPEND   :: 0x00400;
-O_SYNC     :: 0x01000;
-O_ASYNC    :: 0x02000;
-O_CLOEXEC  :: 0x80000;
-
-stdout: Handle;
-stderr: Handle;
-stdin: Handle;
-
-
-write :: proc(fd: Handle, data: []byte) -> (int, Errno) {
-	return 0, 0;
-}
-read :: proc(fd: Handle, data: []byte) -> (int, Errno) {
-	return 0, 0;
-}
-open :: proc(path: string, mode: int = O_RDONLY, perm: int = 0) -> (Handle, Errno) {
-	return 0, 0;
-}
-close :: proc(fd: Handle) -> Errno {
-	return 0;
-}
-seek :: proc(fd: Handle, offset: i64, whence: int) -> (i64, Errno) {
-	return 0, 0;
-}
-current_thread_id :: proc "contextless" () -> int {
-	return 0;
-}
-
-
-file_size :: proc(fd: Handle) -> (i64, Errno) {
-	return 0, 0;
-}
-
-
-
-heap_alloc :: proc(size: int) -> rawptr {
-	return nil;
-}
-heap_resize :: proc(ptr: rawptr, new_size: int) -> rawptr {
-	if new_size == 0 {
-		heap_free(ptr);
-		return nil;
-	}
-	if ptr == nil {
-		return heap_alloc(new_size);
-	}
-
-	return nil;
-}
-heap_free :: proc(ptr: rawptr) {
-	if ptr == nil {
-		return;
-	}
-}

core/os/os_wasi.odin

@@ -0,0 +1,97 @@
+package os
+
+import "core:sys/wasm/wasi"
+
+Handle :: distinct i32
+Errno :: distinct i32
+
+ERROR_NONE :: Errno(wasi.errno_t.SUCCESS)
+
+O_RDONLY   :: 0x00000
+O_WRONLY   :: 0x00001
+O_RDWR     :: 0x00002
+O_CREATE   :: 0x00040
+O_EXCL     :: 0x00080
+O_NOCTTY   :: 0x00100
+O_TRUNC    :: 0x00200
+O_NONBLOCK :: 0x00800
+O_APPEND   :: 0x00400
+O_SYNC     :: 0x01000
+O_ASYNC    :: 0x02000
+O_CLOEXEC  :: 0x80000
+
+stdin:  Handle = 0
+stdout: Handle = 1
+stderr: Handle = 2
+
+
+write :: proc(fd: Handle, data: []byte) -> (int, Errno) {
+	iovs := wasi.ciovec_t(data)
+	n, err := wasi.fd_write(wasi.fd_t(fd), {iovs})
+	return int(n), Errno(err)
+}
+read :: proc(fd: Handle, data: []byte) -> (int, Errno) {
+	iovs := wasi.iovec_t(data)
+	n, err := wasi.fd_read(wasi.fd_t(fd), {iovs})
+	return int(n), Errno(err)
+}
+write_at :: proc(fd: Handle, data: []byte, offset: i64) -> (int, Errno) {
+	iovs := wasi.ciovec_t(data)
+	n, err := wasi.fd_pwrite(wasi.fd_t(fd), {iovs}, wasi.filesize_t(offset))
+	return int(n), Errno(err)
+}
+read_at :: proc(fd: Handle, data: []byte, offset: i64) -> (int, Errno) {
+	iovs := wasi.iovec_t(data)
+	n, err := wasi.fd_pread(wasi.fd_t(fd), {iovs}, wasi.filesize_t(offset))
+	return int(n), Errno(err)
+}
+open :: proc(path: string, mode: int = O_RDONLY, perm: int = 0) -> (Handle, Errno) {
+	return 0, -1
+}
+close :: proc(fd: Handle) -> Errno {
+	err := wasi.fd_close(wasi.fd_t(fd))
+	return Errno(err)
+}
+seek :: proc(fd: Handle, offset: i64, whence: int) -> (i64, Errno) {
+	n, err := wasi.fd_seek(wasi.fd_t(fd), wasi.filedelta_t(offset), wasi.whence_t(whence))
+	return i64(n), Errno(err)
+}
+current_thread_id :: proc "contextless" () -> int {
+	return 0
+}
+
+
+file_size :: proc(fd: Handle) -> (i64, Errno) {
+	stat, err := wasi.fd_filestat_get(wasi.fd_t(fd))
+	if err != nil {
+		return 0, Errno(err)
+	}
+	return i64(stat.size), 0
+}
+
+
+
+heap_alloc :: proc(size: int) -> rawptr {
+	return nil
+}
+heap_resize :: proc(ptr: rawptr, new_size: int) -> rawptr {
+	if new_size == 0 {
+		heap_free(ptr)
+		return nil
+	}
+	if ptr == nil {
+		return heap_alloc(new_size)
+	}
+
+	return nil
+}
+heap_free :: proc(ptr: rawptr) {
+	if ptr == nil {
+		return
+	}
+}
+
+
+exit :: proc "contextless" (code: int) -> ! {
+	wasi.proc_exit(wasi.exitcode_t(code))
+}

core/os/stream.odin

@@ -19,7 +19,7 @@ _file_stream_vtable := &io.Stream_VTable{
 		return
 	},
 	impl_read_at = proc(s: io.Stream, p: []byte, offset: i64) -> (n: int, err: io.Error) {
-		when ODIN_OS == "windows" {
+		when ODIN_OS == "windows" || ODIN_OS == "wasi" {
 			fd := Handle(uintptr(s.stream_data))
 			os_err: Errno
 			n, os_err = read_at(fd, p, offset)
@@ -33,7 +33,7 @@ _file_stream_vtable := &io.Stream_VTable{
 		return
 	},
 	impl_write_at = proc(s: io.Stream, p: []byte, offset: i64) -> (n: int, err: io.Error) {
-		when ODIN_OS == "windows" {
+		when ODIN_OS == "windows" || ODIN_OS == "wasi" {
 			fd := Handle(uintptr(s.stream_data))
 			os_err: Errno
 			n, os_err = write_at(fd, p, offset)

core/runtime/default_allocators_general.odin

@@ -1,5 +1,6 @@
 //+build !windows
 //+build !freestanding
+//+build !wasi
 package runtime
 
 when ODIN_DEFAULT_TO_NIL_ALLOCATOR {

core/runtime/default_allocators_wasi.odin

@@ -0,0 +1,32 @@
+//+build wasi
+package runtime
+
+default_allocator_proc :: proc(allocator_data: rawptr, mode: Allocator_Mode,
+                               size, alignment: int,
+                               old_memory: rawptr, old_size: int, loc := #caller_location) -> ([]byte, Allocator_Error) {
+	switch mode {
+	case .Alloc:
+		return nil, .Out_Of_Memory
+	case .Free:
+		return nil, .None
+	case .Free_All:
+		return nil, .Mode_Not_Implemented
+	case .Resize:
+		if size == 0 {
+			return nil, .None
+		}
+		return nil, .Out_Of_Memory
+	case .Query_Features:
+		return nil, .Mode_Not_Implemented
+	case .Query_Info:
+		return nil, .Mode_Not_Implemented
+	}
+	return nil, .None
+}
+
+default_allocator :: proc() -> Allocator {
+	return Allocator{
+		procedure = default_allocator_proc,
+		data = nil,
+	}
+}

core/runtime/default_temporary_allocator.odin

@@ -1,191 +1,202 @@
 package runtime
 
-@(private)
-byte_slice :: #force_inline proc "contextless" (data: rawptr, len: int) -> []byte {
-	return transmute([]u8)Raw_Slice{data=data, len=max(len, 0)}
-}
-
-
 DEFAULT_TEMP_ALLOCATOR_BACKING_SIZE: int : #config(DEFAULT_TEMP_ALLOCATOR_BACKING_SIZE, 1<<22)
 
 
-Default_Temp_Allocator :: struct {
+when ODIN_OS == "freestanding" {
-	data:               []byte,
+	Default_Temp_Allocator :: struct {
-	curr_offset:        int,
-	prev_allocation:    rawptr,
-	backup_allocator:   Allocator,
-	leaked_allocations: [dynamic][]byte,
-}
-
-default_temp_allocator_init :: proc(s: ^Default_Temp_Allocator, size: int, backup_allocator := context.allocator) {
-	s.data = make_aligned([]byte, size, 2*align_of(rawptr), backup_allocator)
-	s.curr_offset = 0
-	s.prev_allocation = nil
-	s.backup_allocator = backup_allocator
-	s.leaked_allocations.allocator = backup_allocator
-}
-
-default_temp_allocator_destroy :: proc(s: ^Default_Temp_Allocator) {
-	if s == nil {
-		return
 	}
-	for ptr in s.leaked_allocations {
+	
-		free(raw_data(ptr), s.backup_allocator)
+	default_temp_allocator_init :: proc(s: ^Default_Temp_Allocator, size: int, backup_allocator := context.allocator) {
 	}
-	delete(s.leaked_allocations)
+	
-	delete(s.data, s.backup_allocator)
+	default_temp_allocator_destroy :: proc(s: ^Default_Temp_Allocator) {
-	s^ = {}
-}
-
-@(private)
-default_temp_allocator_alloc :: proc(s: ^Default_Temp_Allocator, size, alignment: int, loc := #caller_location) -> ([]byte, Allocator_Error) {
-	size := size
-	size = align_forward_int(size, alignment)
-
-	switch {
-	case s.curr_offset+size <= len(s.data):
-		start := uintptr(raw_data(s.data))
-		ptr := start + uintptr(s.curr_offset)
-		ptr = align_forward_uintptr(ptr, uintptr(alignment))
-		mem_zero(rawptr(ptr), size)
-
-		s.prev_allocation = rawptr(ptr)
-		offset := int(ptr - start)
-		s.curr_offset = offset + size
-		return byte_slice(rawptr(ptr), size), .None
-
-	case size <= len(s.data):
-		start := uintptr(raw_data(s.data))
-		ptr := align_forward_uintptr(start, uintptr(alignment))
-		mem_zero(rawptr(ptr), size)
-
-		s.prev_allocation = rawptr(ptr)
-		offset := int(ptr - start)
-		s.curr_offset = offset + size
-		return byte_slice(rawptr(ptr), size), .None
 	}
-	a := s.backup_allocator
+	
-	if a.procedure == nil {
+	default_temp_allocator_proc :: proc(allocator_data: rawptr, mode: Allocator_Mode,
-		a = context.allocator
+	                                    size, alignment: int,
-		s.backup_allocator = a
+	                                    old_memory: rawptr, old_size: int, loc := #caller_location) -> (data: []byte, err: Allocator_Error) {
+		return nil, nil		
 	}
-
+} else {
-	data, err := mem_alloc_bytes(size, alignment, a, loc)
+	Default_Temp_Allocator :: struct {
-	if err != nil {
+		data:               []byte,
-		return data, err
+		curr_offset:        int,
+		prev_allocation:    rawptr,
+		backup_allocator:   Allocator,
+		leaked_allocations: [dynamic][]byte,
 	}
-	if s.leaked_allocations == nil {
+	
-		s.leaked_allocations = make([dynamic][]byte, a)
+	default_temp_allocator_init :: proc(s: ^Default_Temp_Allocator, size: int, backup_allocator := context.allocator) {
-	}
+		s.data = make_aligned([]byte, size, 2*align_of(rawptr), backup_allocator)
-	append(&s.leaked_allocations, data)
+		s.curr_offset = 0
-
-	// TODO(bill): Should leaks be notified about?
-	if logger := context.logger; logger.lowest_level <= .Warning {
-		if logger.procedure != nil {
-			logger.procedure(logger.data, .Warning, "default temp allocator resorted to backup_allocator" , logger.options, loc)
-		}
-	}
-
-	return data, .None
-}
-
-@(private)
-default_temp_allocator_free :: proc(s: ^Default_Temp_Allocator, old_memory: rawptr, loc := #caller_location) -> Allocator_Error {
-	if old_memory == nil {
-		return .None
-	}
-
-	start := uintptr(raw_data(s.data))
-	end := start + uintptr(len(s.data))
-	old_ptr := uintptr(old_memory)
-
-	if s.prev_allocation == old_memory {
-		s.curr_offset = int(uintptr(s.prev_allocation) - start)
 		s.prev_allocation = nil
-		return .None
+		s.backup_allocator = backup_allocator
+		s.leaked_allocations.allocator = backup_allocator
 	}
 
-	if start <= old_ptr && old_ptr < end {
+	default_temp_allocator_destroy :: proc(s: ^Default_Temp_Allocator) {
-		// NOTE(bill): Cannot free this pointer but it is valid
+		if s == nil {
-		return .None
+			return
+		}
+		for ptr in s.leaked_allocations {
+			free(raw_data(ptr), s.backup_allocator)
+		}
+		delete(s.leaked_allocations)
+		delete(s.data, s.backup_allocator)
+		s^ = {}
 	}
 
-	if len(s.leaked_allocations) != 0 {
+	@(private)
-		for data, i in s.leaked_allocations {
+	default_temp_allocator_alloc :: proc(s: ^Default_Temp_Allocator, size, alignment: int, loc := #caller_location) -> ([]byte, Allocator_Error) {
-			ptr := raw_data(data)
+		size := size
-			if ptr == old_memory {
+		size = align_forward_int(size, alignment)
-				free(ptr, s.backup_allocator)
+
-				ordered_remove(&s.leaked_allocations, i)
+		switch {
-				return .None
+		case s.curr_offset+size <= len(s.data):
+			start := uintptr(raw_data(s.data))
+			ptr := start + uintptr(s.curr_offset)
+			ptr = align_forward_uintptr(ptr, uintptr(alignment))
+			mem_zero(rawptr(ptr), size)
+
+			s.prev_allocation = rawptr(ptr)
+			offset := int(ptr - start)
+			s.curr_offset = offset + size
+			return byte_slice(rawptr(ptr), size), .None
+
+		case size <= len(s.data):
+			start := uintptr(raw_data(s.data))
+			ptr := align_forward_uintptr(start, uintptr(alignment))
+			mem_zero(rawptr(ptr), size)
+
+			s.prev_allocation = rawptr(ptr)
+			offset := int(ptr - start)
+			s.curr_offset = offset + size
+			return byte_slice(rawptr(ptr), size), .None
+		}
+		a := s.backup_allocator
+		if a.procedure == nil {
+			a = context.allocator
+			s.backup_allocator = a
+		}
+
+		data, err := mem_alloc_bytes(size, alignment, a, loc)
+		if err != nil {
+			return data, err
+		}
+		if s.leaked_allocations == nil {
+			s.leaked_allocations = make([dynamic][]byte, a)
+		}
+		append(&s.leaked_allocations, data)
+
+		// TODO(bill): Should leaks be notified about?
+		if logger := context.logger; logger.lowest_level <= .Warning {
+			if logger.procedure != nil {
+				logger.procedure(logger.data, .Warning, "default temp allocator resorted to backup_allocator" , logger.options, loc)
 			}
 		}
-	}
-	return .Invalid_Pointer
-	// panic("invalid pointer passed to default_temp_allocator");
-}
 
-@(private)
+		return data, .None
-default_temp_allocator_free_all :: proc(s: ^Default_Temp_Allocator, loc := #caller_location) {
-	s.curr_offset = 0
-	s.prev_allocation = nil
-	for data in s.leaked_allocations {
-		free(raw_data(data), s.backup_allocator)
-	}
-	clear(&s.leaked_allocations)
-}
-
-@(private)
-default_temp_allocator_resize :: proc(s: ^Default_Temp_Allocator, old_memory: rawptr, old_size, size, alignment: int, loc := #caller_location) -> ([]byte, Allocator_Error) {
-	begin := uintptr(raw_data(s.data))
-	end := begin + uintptr(len(s.data))
-	old_ptr := uintptr(old_memory)
-	if old_memory == s.prev_allocation && old_ptr & uintptr(alignment)-1 == 0 {
-		if old_ptr+uintptr(size) < end {
-			s.curr_offset = int(old_ptr-begin)+size
-			return byte_slice(old_memory, size), .None
-		}
-	}
-	data, err := default_temp_allocator_alloc(s, size, alignment, loc)
-	if err == .None {
-		copy(data, byte_slice(old_memory, old_size))
-		err = default_temp_allocator_free(s, old_memory, loc)
-	}
-	return data, err
-}
-
-default_temp_allocator_proc :: proc(allocator_data: rawptr, mode: Allocator_Mode,
-                                    size, alignment: int,
-                                    old_memory: rawptr, old_size: int, loc := #caller_location) -> (data: []byte, err: Allocator_Error) {
-
-	s := (^Default_Temp_Allocator)(allocator_data)
-
-	if s.data == nil {
-		default_temp_allocator_init(s, DEFAULT_TEMP_ALLOCATOR_BACKING_SIZE, default_allocator())
 	}
 
-	switch mode {
+	@(private)
-	case .Alloc:
+	default_temp_allocator_free :: proc(s: ^Default_Temp_Allocator, old_memory: rawptr, loc := #caller_location) -> Allocator_Error {
-		data, err = default_temp_allocator_alloc(s, size, alignment, loc)
+		if old_memory == nil {
-	case .Free:
+			return .None
-		err = default_temp_allocator_free(s, old_memory, loc)
-
-	case .Free_All:
-		default_temp_allocator_free_all(s, loc)
-
-	case .Resize:
-		data, err = default_temp_allocator_resize(s, old_memory, old_size, size, alignment, loc)
-
-	case .Query_Features:
-		set := (^Allocator_Mode_Set)(old_memory)
-		if set != nil {
-			set^ = {.Alloc, .Free, .Free_All, .Resize, .Query_Features}
 		}
 
-	case .Query_Info:
+		start := uintptr(raw_data(s.data))
-		// Nothing to give
+		end := start + uintptr(len(s.data))
+		old_ptr := uintptr(old_memory)
+
+		if s.prev_allocation == old_memory {
+			s.curr_offset = int(uintptr(s.prev_allocation) - start)
+			s.prev_allocation = nil
+			return .None
+		}
+
+		if start <= old_ptr && old_ptr < end {
+			// NOTE(bill): Cannot free this pointer but it is valid
+			return .None
+		}
+
+		if len(s.leaked_allocations) != 0 {
+			for data, i in s.leaked_allocations {
+				ptr := raw_data(data)
+				if ptr == old_memory {
+					free(ptr, s.backup_allocator)
+					ordered_remove(&s.leaked_allocations, i)
+					return .None
+				}
+			}
+		}
+		return .Invalid_Pointer
+		// panic("invalid pointer passed to default_temp_allocator");
 	}
 
-	return
+	@(private)
+	default_temp_allocator_free_all :: proc(s: ^Default_Temp_Allocator, loc := #caller_location) {
+		s.curr_offset = 0
+		s.prev_allocation = nil
+		for data in s.leaked_allocations {
+			free(raw_data(data), s.backup_allocator)
+		}
+		clear(&s.leaked_allocations)
+	}
+
+	@(private)
+	default_temp_allocator_resize :: proc(s: ^Default_Temp_Allocator, old_memory: rawptr, old_size, size, alignment: int, loc := #caller_location) -> ([]byte, Allocator_Error) {
+		begin := uintptr(raw_data(s.data))
+		end := begin + uintptr(len(s.data))
+		old_ptr := uintptr(old_memory)
+		if old_memory == s.prev_allocation && old_ptr & uintptr(alignment)-1 == 0 {
+			if old_ptr+uintptr(size) < end {
+				s.curr_offset = int(old_ptr-begin)+size
+				return byte_slice(old_memory, size), .None
+			}
+		}
+		data, err := default_temp_allocator_alloc(s, size, alignment, loc)
+		if err == .None {
+			copy(data, byte_slice(old_memory, old_size))
+			err = default_temp_allocator_free(s, old_memory, loc)
+		}
+		return data, err
+	}
+
+	default_temp_allocator_proc :: proc(allocator_data: rawptr, mode: Allocator_Mode,
+	                                    size, alignment: int,
+	                                    old_memory: rawptr, old_size: int, loc := #caller_location) -> (data: []byte, err: Allocator_Error) {
+
+		s := (^Default_Temp_Allocator)(allocator_data)
+
+		if s.data == nil {
+			default_temp_allocator_init(s, DEFAULT_TEMP_ALLOCATOR_BACKING_SIZE, default_allocator())
+		}
+
+		switch mode {
+		case .Alloc:
+			data, err = default_temp_allocator_alloc(s, size, alignment, loc)
+		case .Free:
+			err = default_temp_allocator_free(s, old_memory, loc)
+
+		case .Free_All:
+			default_temp_allocator_free_all(s, loc)
+
+		case .Resize:
+			data, err = default_temp_allocator_resize(s, old_memory, old_size, size, alignment, loc)
+
+		case .Query_Features:
+			set := (^Allocator_Mode_Set)(old_memory)
+			if set != nil {
+				set^ = {.Alloc, .Free, .Free_All, .Resize, .Query_Features}
+			}
+
+		case .Query_Info:
+			// Nothing to give
+		}
+
+		return
+	}
 }
 
 default_temp_allocator :: proc(allocator: ^Default_Temp_Allocator) -> Allocator {
@@ -193,4 +204,4 @@ default_temp_allocator :: proc(allocator: ^Default_Temp_Allocator) -> Allocator
 		procedure = default_temp_allocator_proc,
 		data = allocator,
 	}
-}
+}

core/runtime/internal.odin

@@ -2,6 +2,11 @@ package runtime
 
 import "core:intrinsics"
 
+@(private)
+byte_slice :: #force_inline proc "contextless" (data: rawptr, len: int) -> []byte #no_bounds_check {
+	return ([^]byte)(data)[:max(len, 0)]
+}
+
 bswap_16 :: proc "contextless" (x: u16) -> u16 {
 	return x>>8 | x<<8
 }
@@ -833,3 +838,92 @@ fixunsdfdi :: #force_no_inline proc "c" (a: f64) -> i128 {
 	x := i64(a)
 	return i128(x)
 }
+
+
+
+
+@(link_name="__umodti3")
+umodti3 :: proc "c" (a, b: u128) -> u128 {
+	r: u128 = ---
+	_ = udivmod128(a, b, &r)
+	return r
+}
+
+
+@(link_name="__udivmodti4")
+udivmodti4 :: proc "c" (a, b: u128, rem: ^u128) -> u128 {
+	return udivmod128(a, b, rem)
+}
+
+@(link_name="__udivti3")
+udivti3 :: proc "c" (a, b: u128) -> u128 {
+	return udivmodti4(a, b, nil)
+}
+
+
+@(link_name="__modti3")
+modti3 :: proc "c" (a, b: i128) -> i128 {
+	s_a := a >> (128 - 1)
+	s_b := b >> (128 - 1)
+	an := (a ~ s_a) - s_a
+	bn := (b ~ s_b) - s_b
+
+	r: u128 = ---
+	_ = udivmod128(transmute(u128)an, transmute(u128)bn, &r)
+	return (transmute(i128)r ~ s_a) - s_a
+}
+
+
+@(link_name="__divmodti4")
+divmodti4 :: proc "c" (a, b: i128, rem: ^i128) -> i128 {
+	u := udivmod128(transmute(u128)a, transmute(u128)b, cast(^u128)rem)
+	return transmute(i128)u
+}
+
+@(link_name="__divti3")
+divti3 :: proc "c" (a, b: i128) -> i128 {
+	u := udivmodti4(transmute(u128)a, transmute(u128)b, nil)
+	return transmute(i128)u
+}
+
+
+@(link_name="__fixdfti")
+fixdfti :: proc(a: u64) -> i128 {
+	significandBits :: 52
+	typeWidth       :: (size_of(u64)*8)
+	exponentBits    :: (typeWidth - significandBits - 1)
+	maxExponent     :: ((1 << exponentBits) - 1)
+	exponentBias    :: (maxExponent >> 1)
+
+	implicitBit     :: (u64(1) << significandBits)
+	significandMask :: (implicitBit - 1)
+	signBit         :: (u64(1) << (significandBits + exponentBits))
+	absMask         :: (signBit - 1)
+	exponentMask    :: (absMask ~ significandMask)
+
+	// Break a into sign, exponent, significand
+	aRep := a
+	aAbs := aRep & absMask
+	sign := i128(-1 if aRep & signBit != 0 else 1)
+	exponent := u64((aAbs >> significandBits) - exponentBias)
+	significand := u64((aAbs & significandMask) | implicitBit)
+
+	// If exponent is negative, the result is zero.
+	if exponent < 0 {
+		return 0
+	}
+
+	// If the value is too large for the integer type, saturate.
+	if exponent >= size_of(i128) * 8 {
+		return max(i128) if sign == 1 else min(i128)
+	}
+
+	// If 0 <= exponent < significandBits, right shift to get the result.
+	// Otherwise, shift left.
+	if exponent < significandBits {
+		return sign * i128(significand >> (significandBits - exponent))
+	} else {
+		return sign * (i128(significand) << (exponent - significandBits))
+	}
+
+}

core/runtime/internal_linux.odin

@@ -1,89 +1 @@
 package runtime
-
-import "core:intrinsics"
-
-@(link_name="__umodti3")
-umodti3 :: proc "c" (a, b: u128) -> u128 {
-	r: u128 = ---
-	_ = udivmod128(a, b, &r)
-	return r
-}
-
-
-@(link_name="__udivmodti4")
-udivmodti4 :: proc "c" (a, b: u128, rem: ^u128) -> u128 {
-	return udivmod128(a, b, rem)
-}
-
-@(link_name="__udivti3")
-udivti3 :: proc "c" (a, b: u128) -> u128 {
-	return udivmodti4(a, b, nil)
-}
-
-
-@(link_name="__modti3")
-modti3 :: proc "c" (a, b: i128) -> i128 {
-	s_a := a >> (128 - 1)
-	s_b := b >> (128 - 1)
-	an := (a ~ s_a) - s_a
-	bn := (b ~ s_b) - s_b
-
-	r: u128 = ---
-	_ = udivmod128(transmute(u128)an, transmute(u128)bn, &r)
-	return (transmute(i128)r ~ s_a) - s_a
-}
-
-
-@(link_name="__divmodti4")
-divmodti4 :: proc "c" (a, b: i128, rem: ^i128) -> i128 {
-	u := udivmod128(transmute(u128)a, transmute(u128)b, cast(^u128)rem)
-	return transmute(i128)u
-}
-
-@(link_name="__divti3")
-divti3 :: proc "c" (a, b: i128) -> i128 {
-	u := udivmodti4(transmute(u128)a, transmute(u128)b, nil)
-	return transmute(i128)u
-}
-
-
-@(link_name="__fixdfti")
-fixdfti :: proc(a: u64) -> i128 {
-	significandBits :: 52
-	typeWidth       :: (size_of(u64)*8)
-	exponentBits    :: (typeWidth - significandBits - 1)
-	maxExponent     :: ((1 << exponentBits) - 1)
-	exponentBias    :: (maxExponent >> 1)
-
-	implicitBit     :: (u64(1) << significandBits)
-	significandMask :: (implicitBit - 1)
-	signBit         :: (u64(1) << (significandBits + exponentBits))
-	absMask         :: (signBit - 1)
-	exponentMask    :: (absMask ~ significandMask)
-
-	// Break a into sign, exponent, significand
-	aRep := a
-	aAbs := aRep & absMask
-	sign := i128(-1 if aRep & signBit != 0 else 1)
-	exponent := u64((aAbs >> significandBits) - exponentBias)
-	significand := u64((aAbs & significandMask) | implicitBit)
-
-	// If exponent is negative, the result is zero.
-	if exponent < 0 {
-		return 0
-	}
-
-	// If the value is too large for the integer type, saturate.
-	if exponent >= size_of(i128) * 8 {
-		return max(i128) if sign == 1 else min(i128)
-	}
-
-	// If 0 <= exponent < significandBits, right shift to get the result.
-	// Otherwise, shift left.
-	if exponent < significandBits {
-		return sign * i128(significand >> (significandBits - exponent))
-	} else {
-		return sign * (i128(significand) << (exponent - significandBits))
-	}
-
-}

core/runtime/internal_windows.odin

@@ -1,89 +1 @@
 package runtime
-
-import "core:intrinsics"
-
-@(link_name="__umodti3")
-umodti3 :: proc "c" (a, b: u128) -> u128 {
-	r: u128 = ---
-	_ = udivmod128(a, b, &r)
-	return r
-}
-
-
-@(link_name="__udivmodti4")
-udivmodti4 :: proc "c" (a, b: u128, rem: ^u128) -> u128 {
-	return udivmod128(a, b, rem)
-}
-
-@(link_name="__udivti3")
-udivti3 :: proc "c" (a, b: u128) -> u128 {
-	return udivmodti4(a, b, nil)
-}
-
-
-@(link_name="__modti3")
-modti3 :: proc "c" (a, b: i128) -> i128 {
-	s_a := a >> (128 - 1)
-	s_b := b >> (128 - 1)
-	an := (a ~ s_a) - s_a
-	bn := (b ~ s_b) - s_b
-
-	r: u128 = ---
-	_ = udivmod128(transmute(u128)an, transmute(u128)bn, &r)
-	return (transmute(i128)r ~ s_a) - s_a
-}
-
-
-@(link_name="__divmodti4")
-divmodti4 :: proc "c" (a, b: i128, rem: ^i128) -> i128 {
-	u := udivmod128(transmute(u128)a, transmute(u128)b, cast(^u128)rem)
-	return transmute(i128)u
-}
-
-@(link_name="__divti3")
-divti3 :: proc "c" (a, b: i128) -> i128 {
-	u := udivmodti4(transmute(u128)a, transmute(u128)b, nil)
-	return transmute(i128)u
-}
-
-
-@(link_name="__fixdfti")
-fixdfti :: proc(a: u64) -> i128 {
-	significandBits :: 52
-	typeWidth       :: (size_of(u64)*8)
-	exponentBits    :: (typeWidth - significandBits - 1)
-	maxExponent     :: ((1 << exponentBits) - 1)
-	exponentBias    :: (maxExponent >> 1)
-
-	implicitBit     :: (u64(1) << significandBits)
-	significandMask :: (implicitBit - 1)
-	signBit         :: (u64(1) << (significandBits + exponentBits))
-	absMask         :: (signBit - 1)
-	exponentMask    :: (absMask ~ significandMask)
-
-	// Break a into sign, exponent, significand
-	aRep := a
-	aAbs := aRep & absMask
-	sign := i128(-1 if aRep & signBit != 0 else 1)
-	exponent := (aAbs >> significandBits) - exponentBias
-	significand := (aAbs & significandMask) | implicitBit
-
-	// If exponent is negative, the result is zero.
-	if exponent < 0 {
-		return 0
-	}
-
-	// If the value is too large for the integer type, saturate.
-	if exponent >= size_of(i128) * 8 {
-		return max(i128) if sign == 1 else min(i128)
-	}
-
-	// If 0 <= exponent < significandBits, right shift to get the result.
-	// Otherwise, shift left.
-	if exponent < significandBits {
-		return sign * i128(significand >> (significandBits - exponent))
-	} else {
-		return sign * (i128(significand) << (exponent - significandBits))
-	}
-
-}

core/runtime/os_specific_any.odin

@@ -1,5 +1,4 @@
-//+build !freestanding
+//+build !freestanding !wasi !windows
-//+build !windows
 package runtime
 
 import "core:os"

core/runtime/os_specific_wasi.odin

@@ -0,0 +1,10 @@
+//+build wasi
+package runtime
+
+import "core:sys/wasm/wasi"
+
+_os_write :: proc "contextless" (data: []byte) -> (int, _OS_Errno) {
+	data := (wasi.ciovec_t)(data)
+	n, err := wasi.fd_write(1, {data})
+	return int(n), _OS_Errno(err)
+}

core/runtime/procs.odin

@@ -1,12 +1,39 @@
 package runtime
 
-memset :: proc "c" (ptr: rawptr, val: i32, len: int) -> rawptr {
+when ODIN_ARCH == "wasm32" || ODIN_ARCH == "wasm64" {
-	if ptr != nil && len != 0 {
+	@(link_name="memset")
-		b := byte(val)
+	memset :: proc "c" (ptr: rawptr, val: i32, len: int) -> rawptr {
-		p := ([^]byte)(ptr)
+		if ptr != nil && len != 0 {
-		for i in 0..<len {
+			b := byte(val)
-			p[i] = b
+			p := ([^]byte)(ptr)
+			for i in 0..<len {
+				p[i] = b
+			}
 		}
+		return ptr
+	}
+	
+	@(link_name="memmove")
+	memmove :: proc "c" (dst, src: rawptr, len: int) -> rawptr {
+		if dst != src {
+			d, s := ([^]byte)(dst), ([^]byte)(src)
+			d_end, s_end := d[len:], s[len:]
+			for i := len-1; i >= 0; i -= 1 {
+				d[i] = s[i]
+			}
+		}
+		return dst
+		
+	}
+} else {
+	memset :: proc "c" (ptr: rawptr, val: i32, len: int) -> rawptr {
+		if ptr != nil && len != 0 {
+			b := byte(val)
+			p := ([^]byte)(ptr)
+			for i in 0..<len {
+				p[i] = b
+			}
+		}
+		return ptr
 	}
-	return ptr
 }

core/runtime/procs_wasm32.odin

@@ -0,0 +1,7 @@
+//+build wasm32
+package runtime
+
+@(link_name="__ashlti3")
+__ashlti3 :: proc "c" (a: i64, b: i32) -> i64 {
+	return a
+}

src/build_settings.cpp

@@ -16,6 +16,8 @@ enum TargetOsKind {
 	TargetOs_linux,
 	TargetOs_essence,
 	TargetOs_freebsd,
+	
+	TargetOs_wasi,
 
 	TargetOs_freestanding,
 
@@ -29,6 +31,7 @@ enum TargetArchKind {
 	TargetArch_386,
 	TargetArch_arm64,
 	TargetArch_wasm32,
+	TargetArch_wasm64,
 
 	TargetArch_COUNT,
 };
@@ -49,6 +52,8 @@ String target_os_names[TargetOs_COUNT] = {
 	str_lit("linux"),
 	str_lit("essence"),
 	str_lit("freebsd"),
+	
+	str_lit("wasi"),
 
 	str_lit("freestanding"),
 };
@@ -59,6 +64,7 @@ String target_arch_names[TargetArch_COUNT] = {
 	str_lit("386"),
 	str_lit("arm64"),
 	str_lit("wasm32"),
+	str_lit("wasm64"),
 };
 
 String target_endian_names[TargetEndian_COUNT] = {
@@ -72,6 +78,7 @@ TargetEndianKind target_endians[TargetArch_COUNT] = {
 	TargetEndian_Little,
 	TargetEndian_Little,
 	TargetEndian_Little,
+	TargetEndian_Little,
 };
 
 #ifndef ODIN_VERSION_RAW
@@ -335,6 +342,26 @@ gb_global TargetMetrics target_freestanding_wasm32 = {
 	str_lit(""),
 };
 
+gb_global TargetMetrics target_freestanding_wasm64 = {
+	TargetOs_freestanding,
+	TargetArch_wasm64,
+	8,
+	16,
+	str_lit("wasm64-freestanding-js"),
+	str_lit(""),
+};
+
+gb_global TargetMetrics target_wasi_wasm32 = {
+	TargetOs_wasi,
+	TargetArch_wasm32,
+	4,
+	8,
+	str_lit("wasm32-wasi-js"),
+	str_lit(""),
+};
+
+
+
 
 
 struct NamedTargetMetrics {
@@ -353,6 +380,8 @@ gb_global NamedTargetMetrics named_targets[] = {
 	{ str_lit("freebsd_386"),    &target_freebsd_386    },
 	{ str_lit("freebsd_amd64"),  &target_freebsd_amd64  },
 	{ str_lit("freestanding_wasm32"), &target_freestanding_wasm32 },
+	// { str_lit("freestanding_wasm64"), &target_freestanding_wasm64 },
+	{ str_lit("wasi_wasm32"), &target_wasi_wasm32 },
 };
 
 NamedTargetMetrics *selected_target_metrics;
@@ -458,11 +487,21 @@ bool find_library_collection_path(String name, String *path) {
 }
 
 bool is_arch_wasm(void) {
-	return build_context.metrics.arch == TargetArch_wasm32;
+	switch (build_context.metrics.arch) {
+	case TargetArch_wasm32:
+	case TargetArch_wasm64:
+		return true;
+	}
+	return false;
 }
 
 bool allow_check_foreign_filepath(void) {
-	return build_context.metrics.arch != TargetArch_wasm32;
+	switch (build_context.metrics.arch) {
+	case TargetArch_wasm32:
+	case TargetArch_wasm64:
+		return false;
+	}
+	return true;
 }
 
 
@@ -869,11 +908,24 @@ void init_build_context(TargetMetrics *cross_target) {
 			bc->link_flags = str_lit("-arch arm64 ");
 			break;
 		}
-
+	} else if (is_arch_wasm()) {
-	} else if (bc->metrics.arch == TargetArch_wasm32) {
+		gbString link_flags = gb_string_make(heap_allocator(), " ");
-		bc->link_flags = str_lit("--no-entry --export-table --export-all --allow-undefined ");
+		// link_flags = gb_string_appendc(link_flags, "--export-all ");
+		// link_flags = gb_string_appendc(link_flags, "--export-table ");
+		link_flags = gb_string_appendc(link_flags, "--allow-undefined ");
+		if (bc->metrics.arch == TargetArch_wasm64) {
+			link_flags = gb_string_appendc(link_flags, "-mwas64 ");
+		}
+		if (bc->metrics.os == TargetOs_freestanding) {
+			link_flags = gb_string_appendc(link_flags, "--no-entry ");
+		}
+		
+		bc->link_flags = make_string_c(link_flags);
+		
+		// Disallow on wasm
+		build_context.use_separate_modules = false;
 	} else {
-		gb_printf_err("Compiler Error: Unsupported architecture\n");;
+		gb_printf_err("Compiler Error: Unsupported architecture\n");
 		gb_exit(1);
 	}
 

src/check_decl.cpp

@@ -899,6 +899,10 @@ void check_proc_decl(CheckerContext *ctx, Entity *e, DeclInfo *d) {
 			mutex_unlock(&ctx->info->foreign_mutex);
 		}
 	}
+	
+	if (e->Procedure.link_name.len > 0 ) {
+		e->flags |= EntityFlag_CustomLinkName;
+	}
 }
 
 void check_global_variable_decl(CheckerContext *ctx, Entity *&e, Ast *type_expr, Ast *init_expr) {
@@ -990,6 +994,10 @@ void check_global_variable_decl(CheckerContext *ctx, Entity *&e, Ast *type_expr,
 			string_map_set(fp, key, e);
 		}
 	}
+	
+	if (e->Variable.link_name.len > 0) {
+		e->flags |= EntityFlag_CustomLinkName;
+	}
 
 	if (init_expr == nullptr) {
 		if (type_expr == nullptr) {

src/checker.cpp

@@ -2011,6 +2011,9 @@ void generate_minimum_dependency_set(Checker *c, Entity *start) {
 		str_lit("gnu_h2f_ieee"),
 		str_lit("gnu_f2h_ieee"),
 		str_lit("extendhfsf2"),
+		
+		// WASM Specific
+		str_lit("__ashlti3"),
 	};
 	for (isize i = 0; i < gb_count_of(required_runtime_entities); i++) {
 		force_add_dependency_entity(c, c->info.runtime_package->scope, required_runtime_entities[i]);

src/entity.cpp

@@ -74,9 +74,10 @@ enum EntityFlag : u64 {
 
 	EntityFlag_Test          = 1ull<<30,
 	EntityFlag_Init          = 1ull<<31,
+	
+	EntityFlag_CustomLinkName = 1ull<<40,
 
 	EntityFlag_Overridden    = 1ull<<63,
-
 };
 
 enum EntityState : u32 {

src/llvm_abi.cpp

@@ -1039,6 +1039,75 @@ namespace lbAbiArm64 {
 	}
 }
 
+namespace lbAbiWasm32 {
+	Array<lbArgType> compute_arg_types(LLVMContextRef c, LLVMTypeRef *arg_types, unsigned arg_count);
+	lbArgType compute_return_type(LLVMContextRef c, LLVMTypeRef return_type, bool return_is_defined);
+
+	LB_ABI_INFO(abi_info) {
+		lbFunctionType *ft = gb_alloc_item(permanent_allocator(), lbFunctionType);
+		ft->ctx = c;
+		ft->args = compute_arg_types(c, arg_types, arg_count);
+		ft->ret = compute_return_type(c, return_type, return_is_defined);
+		ft->calling_convention = calling_convention;
+		return ft;
+	}
+
+	lbArgType non_struct(LLVMContextRef c, LLVMTypeRef type, bool is_return) {
+		if (!is_return && type == LLVMIntTypeInContext(c, 128)) {
+			LLVMTypeRef cast_type = LLVMVectorType(LLVMInt64TypeInContext(c), 2);
+			return lb_arg_type_direct(type, cast_type, nullptr, nullptr);
+		}
+		
+		if (!is_return && lb_sizeof(type) > 8) {
+			return lb_arg_type_indirect(type, nullptr);
+		}
+
+		LLVMAttributeRef attr = nullptr;
+		LLVMTypeRef i1 = LLVMInt1TypeInContext(c);
+		if (type == i1) {
+			attr = lb_create_enum_attribute(c, "zeroext");
+		}
+		return lb_arg_type_direct(type, nullptr, nullptr, attr);
+	}
+
+	Array<lbArgType> compute_arg_types(LLVMContextRef c, LLVMTypeRef *arg_types, unsigned arg_count) {
+		auto args = array_make<lbArgType>(heap_allocator(), arg_count);
+
+		for (unsigned i = 0; i < arg_count; i++) {
+			LLVMTypeRef t = arg_types[i];
+			LLVMTypeKind kind = LLVMGetTypeKind(t);
+			i64 sz = lb_sizeof(t);
+			if (kind == LLVMStructTypeKind || kind == LLVMArrayTypeKind) {
+				if (sz == 0) {
+					args[i] = lb_arg_type_ignore(t);
+				} else {
+					args[i] = lb_arg_type_indirect(t, nullptr);
+				}
+			} else {
+				args[i] = non_struct(c, t, false);
+			}
+		}
+		return args;
+	}
+
+	lbArgType compute_return_type(LLVMContextRef c, LLVMTypeRef return_type, bool return_is_defined) {
+		if (!return_is_defined) {
+			return lb_arg_type_direct(LLVMVoidTypeInContext(c));
+		} else if (lb_is_type_kind(return_type, LLVMStructTypeKind) || lb_is_type_kind(return_type, LLVMArrayTypeKind)) {
+			i64 sz = lb_sizeof(return_type);
+			switch (sz) {
+			case 1: return lb_arg_type_direct(return_type, LLVMIntTypeInContext(c, 32), nullptr, nullptr);
+			case 2: return lb_arg_type_direct(return_type, LLVMIntTypeInContext(c, 32), nullptr, nullptr);
+			case 4: return lb_arg_type_direct(return_type, LLVMIntTypeInContext(c, 32), nullptr, nullptr);
+			case 8: return lb_arg_type_direct(return_type, LLVMIntTypeInContext(c, 64), nullptr, nullptr);
+			}
+			LLVMAttributeRef attr = lb_create_enum_attribute_with_type(c, "sret", return_type);
+			return lb_arg_type_indirect(return_type, attr);
+		}
+		return non_struct(c, return_type, true);
+	}
+}
+
 
 LB_ABI_INFO(lb_get_abi_info) {
 	switch (calling_convention) {
@@ -1061,19 +1130,27 @@ LB_ABI_INFO(lb_get_abi_info) {
 		}
 	}
 
-	if (build_context.metrics.arch == TargetArch_amd64) {
+	switch (build_context.metrics.arch) {
+	case TargetArch_amd64:
 		if (build_context.metrics.os == TargetOs_windows) {
 			return lbAbiAmd64Win64::abi_info(c, arg_types, arg_count, return_type, return_is_defined, calling_convention);
 		} else {
 			return lbAbiAmd64SysV::abi_info(c, arg_types, arg_count, return_type, return_is_defined, calling_convention);
 		}
-	} else if (build_context.metrics.arch == TargetArch_386) {
+	case TargetArch_386:
 		return lbAbi386::abi_info(c, arg_types, arg_count, return_type, return_is_defined, calling_convention);
-	} else if (build_context.metrics.arch == TargetArch_arm64) {
+	case TargetArch_arm64:
 		return lbAbiArm64::abi_info(c, arg_types, arg_count, return_type, return_is_defined, calling_convention);
-	} else if (build_context.metrics.arch == TargetArch_wasm32) {
+	case TargetArch_wasm32:
-		return lbAbi386::abi_info(c, arg_types, arg_count, return_type, return_is_defined, calling_convention);
+		// TODO(bill): implement wasm32's ABI correct 
+		// NOTE(bill): this ABI is only an issue for WASI compatibility
+		return lbAbiWasm32::abi_info(c, arg_types, arg_count, return_type, return_is_defined, calling_convention);
+	case TargetArch_wasm64:
+		// TODO(bill): implement wasm64's ABI correct 
+		// NOTE(bill): this ABI is only an issue for WASI compatibility
+		return lbAbiAmd64SysV::abi_info(c, arg_types, arg_count, return_type, return_is_defined, calling_convention);
 	}
+
 	GB_PANIC("Unsupported ABI");
 	return {};
 }

src/llvm_backend.cpp

@@ -771,6 +771,8 @@ lbProcedure *lb_create_main_procedure(lbModule *m, lbProcedure *startup_runtime)
 	Type *results = alloc_type_tuple();
 
 	Type *t_ptr_cstring = alloc_type_pointer(t_cstring);
+	
+	bool call_cleanup = true;
 
 	bool has_args = false;
 	bool is_dll_main = false;
@@ -782,8 +784,12 @@ lbProcedure *lb_create_main_procedure(lbModule *m, lbProcedure *startup_runtime)
 		params->Tuple.variables[0] = alloc_entity_param(nullptr, make_token_ident("hinstDLL"),   t_rawptr, false, true);
 		params->Tuple.variables[1] = alloc_entity_param(nullptr, make_token_ident("fdwReason"),  t_u32,    false, true);
 		params->Tuple.variables[2] = alloc_entity_param(nullptr, make_token_ident("lpReserved"), t_rawptr, false, true);
+		call_cleanup = false;
 	} else if (build_context.metrics.os == TargetOs_windows && build_context.metrics.arch == TargetArch_386) {
 		name = str_lit("mainCRTStartup");
+	} else if (is_arch_wasm()) {
+		name = str_lit("_start");
+		call_cleanup = false;
 	} else {
 		has_args = true;
 		slice_init(&params->Tuple.variables, permanent_allocator(), 2);
@@ -874,8 +880,10 @@ lbProcedure *lb_create_main_procedure(lbModule *m, lbProcedure *startup_runtime)
 	}
 
 	
-	lbValue cleanup_runtime_value = lb_find_runtime_value(m, str_lit("_cleanup_runtime"));
+	if (call_cleanup) {
-	lb_emit_call(p, cleanup_runtime_value, {}, ProcInlining_none, false);
+		lbValue cleanup_runtime_value = lb_find_runtime_value(m, str_lit("_cleanup_runtime"));
+		lb_emit_call(p, cleanup_runtime_value, {}, ProcInlining_none, false);
+	}
 	
 
 	if (is_dll_main) {
@@ -885,6 +893,19 @@ lbProcedure *lb_create_main_procedure(lbModule *m, lbProcedure *startup_runtime)
 	}
 
 	lb_end_procedure_body(p);
+	
+
+	if (is_arch_wasm()) {
+		LLVMSetLinkage(p->value, LLVMDLLExportLinkage);
+		LLVMSetDLLStorageClass(p->value, LLVMDLLExportStorageClass);
+		LLVMSetVisibility(p->value, LLVMDefaultVisibility);
+		
+		char const *export_name = alloc_cstring(permanent_allocator(), p->name);
+		LLVMAddTargetDependentFunctionAttr(p->value, "wasm-export-name", export_name);
+	} else {
+		LLVMSetLinkage(p->value, LLVMExternalLinkage);
+	}
+	
 
 	if (!m->debug_builder && LLVMVerifyFunction(p->value, LLVMReturnStatusAction)) {
 		gb_printf_err("LLVM CODE GEN FAILED FOR PROCEDURE: %s\n", "main");
@@ -1064,14 +1085,10 @@ struct lbLLVMModulePassWorkerData {
 };
 
 WORKER_TASK_PROC(lb_llvm_module_pass_worker_proc) {
-	GB_ASSERT(MULTITHREAD_OBJECT_GENERATION);
-
 	auto wd = cast(lbLLVMModulePassWorkerData *)data;
-
 	LLVMPassManagerRef module_pass_manager = LLVMCreatePassManager();
 	lb_populate_module_pass_manager(wd->target_machine, module_pass_manager, build_context.optimization_level);
 	LLVMRunPassManager(module_pass_manager, wd->m->mod);
-
 	return 0;
 }
 
@@ -1148,6 +1165,7 @@ void lb_generate_code(lbGenerator *gen) {
 		LLVMInitializeAArch64Disassembler();
 		break;
 	case TargetArch_wasm32:
+	case TargetArch_wasm64:
 		LLVMInitializeWebAssemblyTargetInfo();
 		LLVMInitializeWebAssemblyTarget();
 		LLVMInitializeWebAssemblyTargetMC();
@@ -1660,6 +1678,8 @@ void lb_generate_code(lbGenerator *gen) {
 
 	for_array(i, gen->modules.entries) {
 		lbModule *m = gen->modules.entries[i].value;
+		
+		lb_run_remove_unused_function_pass(m->mod);
 
 		auto wd = gb_alloc_item(permanent_allocator(), lbLLVMModulePassWorkerData);
 		wd->m = m;
@@ -1737,8 +1757,16 @@ void lb_generate_code(lbGenerator *gen) {
 	}
 
 	TIME_SECTION("LLVM Object Generation");
+	
+	isize non_empty_module_count = 0;
+	for_array(j, gen->modules.entries) {
+		lbModule *m = gen->modules.entries[j].value;
+		if (!lb_is_module_empty(m)) {
+			non_empty_module_count += 1;
+		}
+	}
 
-	if (do_threading) {
+	if (do_threading && non_empty_module_count > 1) {
 		for_array(j, gen->modules.entries) {
 			lbModule *m = gen->modules.entries[j].value;
 			if (lb_is_module_empty(m)) {

src/llvm_backend.hpp

@@ -585,3 +585,24 @@ enum : LLVMAttributeIndex {
 	LLVMAttributeIndex_FunctionIndex = ~0u,
 	LLVMAttributeIndex_FirstArgIndex = 1,
 };
+
+
+char const *llvm_linkage_strings[] = {
+	"external linkage",
+	"available externally linkage",
+	"link once any linkage",
+	"link once odr linkage",
+	"link once odr auto hide linkage",
+	"weak any linkage",
+	"weak odr linkage",
+	"appending linkage",
+	"internal linkage",
+	"private linkage",
+	"dllimport linkage",
+	"dllexport linkage",
+	"external weak linkage",
+	"ghost linkage",
+	"common linkage",
+	"linker private linkage",
+	"linker private weak linkage"
+};

src/llvm_backend_expr.cpp

@@ -496,6 +496,7 @@ bool lb_is_matrix_simdable(Type *t) {
 		break;
 	case TargetArch_386:
 	case TargetArch_wasm32:
+	case TargetArch_wasm64:
 		// nope
 		return false;
 	}
@@ -513,6 +514,7 @@ bool lb_is_matrix_simdable(Type *t) {
 				return true;
 			case TargetArch_386:
 			case TargetArch_wasm32:
+			case TargetArch_wasm64:
 				return false;
 			}
 		}

src/llvm_backend_opt.cpp

@@ -355,3 +355,63 @@ void lb_run_function_pass_manager(LLVMPassManagerRef fpm, lbProcedure *p) {
 	// are not removed
 	lb_run_remove_dead_instruction_pass(p);
 }
+
+
+void lb_run_remove_unused_function_pass(LLVMModuleRef mod) {
+	isize removal_count = 0;
+	isize pass_count = 0;
+	isize const max_pass_count = 10;
+	// Custom remove dead function pass
+	for (; pass_count < max_pass_count; pass_count++) {
+		bool was_dead_function = false;	
+		for (LLVMValueRef func = LLVMGetFirstFunction(mod);
+		     func != nullptr;
+		     /**/
+		     ) {
+		     	LLVMValueRef curr_func = func;
+		     	func = LLVMGetNextFunction(func);
+		     	
+			LLVMUseRef first_use = LLVMGetFirstUse(curr_func);
+			if (first_use != nullptr)  {
+				continue;
+			}
+			String name = {};
+			name.text = cast(u8 *)LLVMGetValueName2(curr_func, cast(size_t *)&name.len);
+						
+			if (LLVMIsDeclaration(curr_func)) {
+				// Ignore for the time being
+				continue;
+			}
+			
+			if (name == "memset" ||
+			    name == "memmove" ||
+			    name == "memcpy") {
+				continue;
+			}
+			if (is_arch_wasm()) {
+				if (name == "__ashlti3") {
+					LLVMSetLinkage(curr_func, LLVMExternalLinkage);
+					continue;
+				}
+			}
+			
+			LLVMLinkage linkage = LLVMGetLinkage(curr_func);
+						
+			switch (linkage) {
+			case LLVMExternalLinkage:
+			case LLVMDLLImportLinkage:
+			case LLVMDLLExportLinkage:
+			default:
+				continue;
+			case LLVMInternalLinkage:
+				break;
+			}
+			LLVMDeleteFunction(curr_func);
+			was_dead_function = true;
+			removal_count += 1;
+		}
+		if (!was_dead_function) {
+			break;
+		}
+	}
+}

src/llvm_backend_proc.cpp

@@ -195,13 +195,19 @@ lbProcedure *lb_create_procedure(lbModule *m, Entity *entity, bool ignore_body)
 			// then it is very likely it is required by LLVM and thus cannot have internal linkage
 			if (entity->pkg != nullptr && entity->pkg->kind == Package_Runtime && p->body != nullptr) {
 				GB_ASSERT(entity->kind == Entity_Procedure);
-				if (entity->Procedure.link_name != "") {
+				String link_name = entity->Procedure.link_name;
-					LLVMSetLinkage(p->value, LLVMExternalLinkage);
+				if (entity->flags & EntityFlag_CustomLinkName && 
+				    link_name != "") {
+					if (string_starts_with(link_name, str_lit("__"))) {
+						LLVMSetLinkage(p->value, LLVMExternalLinkage);
+					} else {
+						LLVMSetLinkage(p->value, LLVMInternalLinkage);
+					}
 				}
 			}
 		}
 	}
-
+	
 	if (p->is_foreign) {
 		if (is_arch_wasm()) {
 			char const *import_name = alloc_cstring(permanent_allocator(), p->name);
@@ -217,6 +223,7 @@ lbProcedure *lb_create_procedure(lbModule *m, Entity *entity, bool ignore_body)
 			LLVMAddTargetDependentFunctionAttr(p->value, "wasm-import-module", module_name);
 		}
 	}
+	
 
 	// NOTE(bill): offset==0 is the return value
 	isize offset = 1;

src/llvm_backend_utility.cpp

@@ -1504,6 +1504,7 @@ lbValue lb_emit_mul_add(lbProcedure *p, lbValue a, lbValue b, lbValue c, Type *t
 			break;
 		case TargetArch_386:
 		case TargetArch_wasm32:
+		case TargetArch_wasm64:
 			is_possible = false;
 			break;
 		}

src/main.cpp

@@ -135,13 +135,12 @@ i32 linker_stage(lbGenerator *gen) {
 
 	if (is_arch_wasm()) {
 		timings_start_section(timings, str_lit("wasm-ld"));
+
 		result = system_exec_command_line_app("wasm-ld",
-			"\"%.*s\\bin\\wasm-ld\" \"%.*s.wasm-obj\" -o \"%.*s.wasm\" %.*s %.*s",
+			"\"%.*s\\bin\\wasm-ld\" \"%.*s.wasm.o\" -o \"%.*s.wasm\" %.*s %.*s",
 			LIT(build_context.ODIN_ROOT),
 			LIT(output_base), LIT(output_base), LIT(build_context.link_flags), LIT(build_context.extra_linker_flags));
-		if (result) {
+		return result;
-			return result;
-		}
 	}
 
 	if (build_context.cross_compiling && selected_target_metrics->metrics == &target_essence_amd64) {

src/string.cpp

@@ -21,13 +21,14 @@ struct String {
 };
 // NOTE(bill): used for printf style arguments
 #define LIT(x) ((int)(x).len), (x).text
-#define STR_LIT(c_str) {cast(u8 *)c_str, gb_size_of(c_str)-1}
 #if defined(GB_COMPILER_MSVC) && _MSC_VER < 1700
-	#define str_lit(c_str) make_string(cast(u8 *)c_str, gb_size_of(c_str)-1)
+	#define STR_LIT(c_str) make_string(cast(u8 *)c_str, gb_size_of(c_str)-1)
 #else
-	#define str_lit(c_str) String{cast(u8 *)c_str, gb_size_of(c_str)-1}
+	#define STR_LIT(c_str) String{cast(u8 *)c_str, gb_size_of(c_str)-1}
 #endif
 
+#define str_lit(c_str) STR_LIT(c_str)
+
 // NOTE(bill): String16 is only used for Windows due to its file directories
 struct String16 {
 	wchar_t *text;