Creating shared WebAssembly library

[nomeata]

I am trying to build a shared WebAssembly library (in the sense of https://github.com/WebAssembly/tool-conventions/blob/master/DynamicLinking.md) from Rust.

The following looks promising:

#![no_main]
#![feature(link_args)]
#![allow(unused_attributes)]
#![link_args = "--import-memory"]
#![link_args = "--shared"]

// no_std just to make the output smaller and easier to understand
#![no_std]
#![feature(lang_items)]
#[lang = "eh_personality"] extern fn eh_personality() {}
use core::panic::PanicInfo;
#[panic_handler]
fn panic(_info: &PanicInfo) -> ! { loop {} }


#[link(wasm_import_module = "import")]
extern { pub fn foo(ptr : *const u8); }


#[export_name = "start"]
pub unsafe fn start() {
    foo("Hi".as_ptr());
}

If I compile this (using --target wasm32-unknown-unknown), I get the following code:

(module
 (type $0 (func (param i32)))
 (type $1 (func))
 (type $2 (func (param i32 i32) (result i32)))
 (import "env" "memory" (memory $0 0))
 (import "env" "__indirect_function_table" (table 0 anyfunc))
 (import "env" "__stack_pointer" (global $gimport$2 i32))
 (import "env" "__memory_base" (global $gimport$3 i32))
 (import "env" "__table_base" (global $gimport$4 i32))
 (import "import" "foo" (func $foo (param i32)))
 (global $global$0 i32 (i32.const 0))
 (data (get_global $gimport$3) "\01gdb_load_rust_pretty_printers.py\00Hi")
 (export "__rustc_debug_gdb_scripts_section__" (global $global$0))
 (export "start" (func $start))
 (func $__wasm_call_ctors (; 1 ;) (type $1)
 )
 (func $_ZN4core3str21_$LT$impl$u20$str$GT$6as_ptr17hd1ab9e60453863b8E (; 2 ;) (type $2) (param $var$0 i32) (param $var$1 i32) (result i32)
  (local $var$2 i32)
  (local $var$3 i32)
  (local $var$4 i32)
  (local $var$5 i32)
  (set_local $var$2
   (get_global $gimport$2)
  )
  (set_local $var$3
   (i32.const 16)
  )
  (set_local $var$4
   (i32.sub
    (get_local $var$2)
    (get_local $var$3)
   )
  )
  (i32.store offset=8
   (get_local $var$4)
   (get_local $var$0)
  )
  (i32.store offset=12
   (get_local $var$4)
   (get_local $var$1)
  )
  (set_local $var$5
   (i32.load offset=8
    (get_local $var$4)
   )
  )
  (return
   (get_local $var$5)
  )
 )
 (func $start (; 3 ;) (type $1)
  (local $var$0 i32)
  (local $var$1 i32)
  (local $var$2 i32)
  (local $var$3 i32)
  (set_local $var$0
   (i32.const 34)
  )
  (set_local $var$1
   (get_local $var$0)
  )
  (set_local $var$2
   (i32.const 2)
  )
  (set_local $var$3
   (call $_ZN4core3str21_$LT$impl$u20$str$GT$6as_ptr17hd1ab9e60453863b8E
    (get_local $var$1)
    (get_local $var$2)
   )
  )
  (call $foo
   (get_local $var$3)
  )
  (return)
 )
 ;; custom section "dylink", size 5
 ;; custom section ".debug_info", size 1748
 ;; custom section ".debug_macinfo", size 2
 ;; custom section ".debug_pubtypes", size 959
 ;; custom section ".debug_ranges", size 32
 ;; custom section ".debug_abbrev", size 401
 ;; custom section ".debug_line", size 280
 ;; custom section ".debug_str", size 1569
 ;; custom section ".debug_pubnames", size 235
 ;; custom section "producers", size 79
)

It is very promising that this module imports __stack_pointer and __memory_base, and that it puts the static data ("\01gdb_load_rust_pretty_printers.py\00Hi") at a position that is determined by the imported __memory_base.

But when the code actually references the Hi string, in function start, it simply uses an absolute pointer, without shifting it by __memory_base:

  (set_local $var$0
   (i32.const 34)
  )

so this does not seem to be working yet.

Is there a flag I am using wrong? Or is this just not yet supported?

$ rustc --version
rustc 1.35.0-nightly (3750348da 2019-04-08)

[jonas-schievink]

Why does this need to be done? Locals should always refer to the function's stack frame implicitly.

[nomeata]

Why does this need to be done? Locals should always refer to the function's stack frame implicitly.

I am not referring to locals; the problem is the (i32.const 34): This seems to be the pointer to "Hi", as an absolute pointer. I would expect this to be (i32.add (get_global $gimport$2) (i32.const 34)), so that the address is taken relative to the __memory_base.

[nomeata]

JFTR, same happens with crate-type = ["cdylib"]

[jonas-schievink]

Ah I see, that makes more sense

[nomeata]

To reproduce without cargo, I am using this command line:

rustc  foo.rs  --crate-type cdylib -o foo.wasm  --target wasm32-unknown-unknown

[nomeata]

With -C incremental=foo I can get my hold on the .o files passed to lld, and these files are WebAssembly modules with a static linking section, including a code relocation section according to https://github.com/WebAssembly/tool-conventions/blob/master/Linking.md, but I can see that lld isn’t really in a position to turn this into a dynamically linkable library.

Relevant bits:

$ wasm-dis foo/lib-1oxx6vmtp942d/s-fbld28zq8j-122tlgg-2xj38ws0mwqgo/x73358tcv87tyh4.o
(module
 (type $0 (func))
 (type $1 (func (param i32)))
 (type $2 (func (param i32 i32) (result i32)))
 (import "env" "__linear_memory" (memory $0 1))
 (import "env" "__indirect_function_table" (table 0 anyfunc))
 (import "env" "_ZN4core3str21_$LT$impl$u20$str$GT$6as_ptr17h490791c777cb2b36E" (func $fimport$2 (param i32 i32) (result i32)))
 (import "import" "foo" (func $fimport$3 (param i32)))
 (data (i32.const 0) "Hi")
…
 (func $2 (; 4 ;) (type $0)
…
  (set_local $var$0
   (i32.const 0)
  )
…
 )
 ;; custom section "linking", size 117
 ;; custom section "reloc.CODE", size 12
)

[nomeata]

rustc is using LLVM to generate and link WebAssembly code, right? Just to narrow down what’s happening where, here is the equivalent in C:

/tmp $ cat foo.c 
extern int bar (int, char *);

char *s = "Foo";

int foo (int n)  { return bar(n, s); }
/tmp $ clang-9 --compile -fpic foo.c  --target=wasm32-unknown-unknown-wasm --optimize=3 --output foo.o
/tmp $ wasm-ld-9 --shared --no-entry foo.o -o foo.wasm --export=foo --gc-sections
/tmp $ wasm-dis foo.wasm 
(module
 (type $0 (func (param i32 i32) (result i32)))
 (type $1 (func))
 (type $2 (func (param i32) (result i32)))
 (import "env" "memory" (memory $0 0))
 (import "env" "__indirect_function_table" (table 0 anyfunc))
 (import "env" "__memory_base" (global $gimport$2 i32))
 (import "env" "__table_base" (global $gimport$3 i32))
 (import "env" "bar" (func $bar (param i32 i32) (result i32)))
 (data (get_global $gimport$2) "Foo\00\00\00\00\00")
 (export "foo" (func $foo))
 (func $__wasm_call_ctors (; 1 ;) (type $1)
  (call $__wasm_apply_relocs)
 )
 (func $__wasm_apply_relocs (; 2 ;) (type $1)
  (i32.store
   (i32.add
    (get_global $gimport$2)
    (i32.const 4)
   )
   (i32.add
    (get_global $gimport$2)
    (i32.const 0)
   )
  )
 )
 (func $foo (; 3 ;) (type $2) (param $var$0 i32) (result i32)
  (call $bar
   (get_local $var$0)
   (i32.load
    (i32.add
     (get_global $gimport$2)
     (i32.const 4)
    )
   )
  )
 )
 ;; custom section "dylink", size 5
 ;; custom section "producers", size 52
)

Note how $foo calculates the address to the static memory based on $gimport$2.

[nomeata]

Since I passed -fpic to clang here, I tried passing -C relocation-model=pic to rustc, but the output is unchanged.

[nomeata]

Possibly related:

• Panic when used on .wasm built with -C link-arg=--shared rustwasm/wasm-bindgen#1420 (comment)
• wasm32: Default to a "static" relocation model #59712
  Maybe the conclusion is that rustc simply doesn’t support producing the PIC code needed for shared libraries.

[Arnavion]

FWIW, back in #46645 (comment) (March 2018) I had success with making a relocatable WASM module. Of course it requires a preprocessing bundler / the WASM loader to actually relocate the modules, so it's not as trivial as a PIC module.

[nomeata]

For my use case, I decided that PIC (e.g. importing a global that indicates memory offeset) is more suited than static relocation sections.

[RReverser]

@alexcrichton It does look like #59712 is related.

At least passing -C relocation-model=pic in addition to -C link-arg=--shared doesn't seem to have any effect, so addresses stay incorrect (static).

Do you think there's a way to override relocation mode back to PIC for consumer experimentation with shared Wasm libraries?

[alexcrichton]

AFAIK no work really has been done to get Rust/wasm working with the dynamic linking proposal. It's pretty highly unlikely that no changes to rustc will be necessary to get things working. I don't personally have time to work on it right now.

My guess is that either the upstream proposal will need to change to support a more ELF-like way of compiling shared objects and linking them into a static-like version, a new target will need to be added, or the standard library will need to be recompiled from scratch with xargo/cargo-xbuild and new codegen flags.

[deprilula28]

Any progress on this?

[osa1]

I tried this with rustc 1.46.0-nightly (0c03aee 2020-07-05) just to see if anything changed in the meantime. Currently -Clink-arg=--shared causes this error:

error: linking with `rust-lld` failed: exit code: 1
  |
  = note: "rust-lld" "-flavor" "wasm" "--no-threads" "-z" "stack-size=1048576" "--stack-first" "--allow-undefined" "--fatal-warnings" "--no-demangle" "--export-dynamic" "--no-entry" "-L" "/home/omer/.rustup/toolchains/nightly-x86_64-unknown-linux-gnu/lib/rustlib/wasm32-unknown-unknown/lib" "-L" "/home/omer/.rustup/toolchains/nightly-x86_64-unknown-linux-gnu/lib/rustlib/wasm32-unknown-unknown/lib/self-contained" "foo.main.7rcbfp3g-cgu.0.rcgu.o" "foo.main.7rcbfp3g-cgu.1.rcgu.o" "-o" "foo.wasm" "--export" "start" "--export=__heap_base" "--export=__data_end" "--gc-sections" "-O0" "-L"
"/home/omer/.rustup/toolchains/nightly-x86_64-unknown-linux-gnu/lib/rustlib/wasm32-unknown-unknown/lib" "/home/omer/.rustup/toolchains/nightly-x86_64-unknown-linux-gnu/lib/rustlib/wasm32-unknown-unknown/lib/librustc_std_workspace_core-338ea5f3d59665eb.rlib" "/home/omer/.rustup/toolchains/nightly-x86_64-unknown-linux-gnu/lib/rustlib/wasm32-unknown-unknown/lib/libcore-c883bdbeb473297b.rlib" "/home/omer/.rustup/toolchains/nightly-x86_64-unknown-linux-gnu/lib/rustlib/wasm32-unknown-unknown/lib/libcompiler_builtins-4a8deb6b3def81fd.rlib" "--shared" "--import-memory"
  = note: rust-lld: error: foo.main.7rcbfp3g-cgu.1.rcgu.o: relocation R_WASM_MEMORY_ADDR_SLEB cannot be used against symbol .L__unnamed_1; recompile with -fPIC


error: aborting due to previous error

If I remove --shared but keep --import-memory I get the same absolute pointer in start. -C relocation-model=... parameters don't make any difference.

[osa1]

Interestingly if I try the C example above:

extern int bar(int, char *);

char *s = "Foo";

int foo(int n)
{
    return bar(n, s);
}

The error is very similar to the one I get from rustc when I try with -Clink-arg=--shared:

$ clang-10 -c -fPIC test.c --target=wasm32-unknown-unknown-wasm -o test.o -O3
$ wasm-ld --shared --no-entry test.o -o test.wasm --export=foo --gc-sections
wasm-ld: error: test.o: relocation R_WASM_MEMORY_ADDR_LEB cannot be used against symbol s; recompile with -fPIC

This says R_WASM_MEMORY_ADDR_LEB instead of R_WASM_MEMORY_ADDR_SLEB but I suspect the root cause is the same.

[osa1]

The lld documentation (https://lld.llvm.org/WebAssembly.html) says

No support for creating shared libraries. The spec for shared libraries in WebAssembly is still in flux: https://github.com/WebAssembly/tool-conventions/blob/master/DynamicLinking.md

So I'm guessing this is something to fix/implement on lld side first before making any changes in rustc?

[osa1]

Regarding the R_WASM_MEMORY_ADDR_SLEB error I mention above, I was reading about this and came across this blog post, which says that the error only happens when using static strings. I confirmed that that's also the case with Rust. For example, in the original reproducer, if I change main:

#![no_main]
#![no_std]
#![feature(link_args, lang_items, core_intrinsics)]
#![allow(unused_attributes)]
#![link_args = "--import-memory"]
#![link_args = "--shared"]

#[panic_handler]
#[no_mangle]
pub fn panic(_info: &::core::panic::PanicInfo) -> ! {
    ::core::intrinsics::abort();
}

#[lang = "eh_personality"]
extern fn rust_eh_personality() {}

#[link(wasm_import_module = "import")]
extern { pub fn foo(ptr : *const u8); }

#[export_name = "main"]
pub unsafe fn main() {
    let x = 5u8;
    foo(&x as *const u8);

    let hi: [u8; 2] = [b'h', b'i'];
    foo(hi.as_ptr());
}

I can build this just fine with

$ rustc lib.rs --target=wasm32-unknown-unknown -O

The wat:

(module
  (type (;0;) (func (param i32)))
  (type (;1;) (func))
  (import "env" "memory" (memory (;0;) 0))
  (import "env" "__indirect_function_table" (table (;0;) 0 funcref))
  (import "env" "__stack_pointer" (global (;0;) (mut i32)))
  (import "env" "__memory_base" (global (;1;) i32))
  (import "env" "__table_base" (global (;2;) i32))
  (import "import" "foo" (func $_ZN3lib3foo17hb22c31d0714767f1E (type 0)))
  (func $__wasm_call_ctors (type 1)
    call $__wasm_apply_relocs)
  (func $__wasm_apply_relocs (type 1))
  (func $main (type 1)
    (local i32)
    global.get 0
    i32.const 16
    i32.sub
    local.tee 0
    global.set 0
    local.get 0
    i32.const 5
    i32.store8 offset=13
    local.get 0
    i32.const 13
    i32.add
    call $_ZN3lib3foo17hb22c31d0714767f1E
    local.get 0
    i32.const 26984
    i32.store16 offset=14 align=1
    local.get 0
    i32.const 14
    i32.add
    call $_ZN3lib3foo17hb22c31d0714767f1E
    local.get 0
    i32.const 16
    i32.add
    global.set 0)
  (export "main" (func $main)))

However if I make the array static

static HI: [u8; 2] = [b'h', b'i'];

#[export_name = "main"]
pub unsafe fn main() {
    let x = 5u8;
    foo(&x as *const u8);
    foo(HI.as_ptr());
}

This fails with the same error.

So it seems like there's something to be fixed in handling of static data. The blog post linked above suggests that this is a clang/LLVM bug as they reproduce the problem in a C file.

[nomeata]

Certianly not a clang/LLVM bug. The blog post refers to LLVM-8 (where indeed it didn't work), but since LLVM-9 I can reliably produce shared WebAssembly libraries from C, including static data.

And it’s not surpring that this is related to static data: If you don’t have static data, there is nothing to relocate upon linking :-)

[osa1]

I'm not sure. Here's an example in C:

extern void f(char*);

static char *s = "blah";

void test(void) {
    f(s);
}

Using clang-10:

$ clang --version
clang version 10.0.0
Target: x86_64-unknown-linux-gnu
Thread model: posix
InstalledDir: /home/omer/Downloads/clang+llvm-10.0.0-x86_64-linux-gnu-ubuntu-18.04/bin

$ clang --target=wasm32-unknown-unknown -nostdlib -c -fPIC -o lib.o lib.c

$ wasm-ld --version
LLD 10.0.0

$ wasm-ld --shared --import-memory --export-all --allow-undefined lib.o -o lib.wasm
wasm-ld: error: lib.o: relocation R_WASM_MEMORY_ADDR_LEB cannot be used against symbol s; recompile with -fPIC

It's not the exact same error though -- it complains about a R_WASM_MEMORY_ADDR_LEB relocation instead of R_WASM_MEMORY_ADDR_SLEB.

[osa1]

I just realized that if I use --target=wasm32-unknown-emscripten above (instead of wasm32-unknown-unknown) it builds:

$ clang --target=wasm32-unknown-emscripten -nostdlib -c -fPIC -o lib.o lib.c
$ wasm-ld --shared --import-memory --export-all --allow-undefined lib.o -o lib.wasm

[osa1]

When building Rust simply using the target wasm32-unknown-emscripten is not enough, we also need --crate-type=lib and -Crelocation-model=pic. Full example:

$ cat lib.rs
#![no_std]
#![feature(lang_items, core_intrinsics)]

#[panic_handler]
#[no_mangle]
pub fn panic(_info: &core::panic::PanicInfo) -> ! {
    core::intrinsics::abort();
}

#[lang = "eh_personality"]
extern "C" fn rust_eh_personality() {}

extern "C" {
    pub fn foo(ptr: *const u8);
}

#[export_name = "_start"]
pub unsafe fn main() {
    foo("hi".as_ptr());
}

$ rustc lib.rs --target=wasm32-unknown-emscripten -O --crate-type=lib -Crelocation-model=pic

This generates a liblib.rlib file which we need to unpack to get the .o file:

$ ar -x liblib.rlib

Now if we look at the code in the .o file:

$ wasm-objdump -d lib.lib.3a1fbbbh-cgu.0.rcgu.o

lib.lib.3a1fbbbh-cgu.0.rcgu.o:  file format wasm 0x1

Code Disassembly:

00008e func[1] <rust_begin_unwind>:
 00008f: 00                         | unreachable
 000090: 00                         | unreachable
 000091: 0b                         | end
000093 func[2] <rust_eh_personality>:
 000094: 0b                         | end
000096 func[3] <_start>:
 000097: 23 80 80 80 80 00          | global.get 0 <env.__memory_base>
 00009d: 41 80 80 80 80 00          | i32.const 0
 0000a3: 6a                         | i32.add
 0000a4: 10 80 80 80 80 00          | call 0 <env.foo>
 0000aa: 0b                         | end

Here the address of the string (static data) is relative to env.__memory_base, which is what we were trying to generate.

[nomeata]

Great, that’s all I wanted :-)

[dalcde]

I'm having difficulties building a cdylib with these instructions. I am able to produce a lib, but when I set crate-type=cdylib, I get wasm-ld: error: unknown file type: lib.rmeta

I get something reasonable-looking if I manually extract the .rlib file and run wasm-ld by hand, but that is intractable for larger projects with lots of dependencies.

[RReverser]

This generates a liblib.rlib file which we need to unpack to get the .o file:

FWIW instead of this, you can ask Rust to produce the object file directly with --emit=obj.

[alienself]

@nomeata @osa1 or anyone on this thread:

• how do I transform the liblib.rlib into a wasm file?
• how can I load the compiled shared library from another wasm file and bind those two using javascript?