Miscompiled program with `--target=wasm32`

[zack-alex]

UPD: here is a slightly improved example.

Here is a simple C++ program:

struct BigClass {
  char data[400000];
};

struct Class {
  char big_data[400000];
  int value{33};
};

Class glob{};

extern "C" int _start() {
  glob.value = 42;
  BigClass s1{};
  return glob.value;
}

When compiled and ran with wasmer runtime, it outputs "0" instead of expected "42":

+ clang-17 --version
clang version 17.0.6
Target: aarch64-apple-darwin23.6.0
Thread model: posix
InstalledDir: /nix/store/glzjy0xq2xhmwg6pc8c1pdk7gln127wj-clang-17.0.6/bin
+ clang-17 -std=c++20 -pedantic -Wall -Wno-unused-variable --target=wasm32 --no-standard-libraries -mbulk-memory -O0 -o main.wasm main.cpp
+ wasmer main.wasm
0

[llvmbot]

@llvm/issue-subscribers-backend-webassembly

Author: Zach Philipchook (tekhnus)

Here is a simple C++ program: ``` struct BigClass { char data[400000]; };

struct Class {
char big_data[400000];
int value{33};
};

Class glob{};

extern "C" int _start() {
glob.value = 42;
BigClass s1{};
return glob.value;
}

When compiled and ran with `wasmer` runtime, it outputs "0" instead of expected "42":

• clang-17 --version
  clang version 17.0.6
  Target: aarch64-apple-darwin23.6.0
  Thread model: posix
  InstalledDir: /nix/store/glzjy0xq2xhmwg6pc8c1pdk7gln127wj-clang-17.0.6/bin
• clang-17 -std=c++20 -pedantic -Wall -Wno-unused-variable --target=wasm32 --no-standard-libraries -mbulk-memory -O0 -o main.wasm main.cpp
• wasmer main.wasm
  0

</details>

[sbc100]

There are a couple of issues with this program. Firstly, the _start function does not return a value, it returns void. If you want to return a value you should probably call __wasi_proc_exit. e.g. __wasi_proc_exit(glob.value);

Secondly, by implementing _start rather then main here you are bypassing the normal C startup stuff such as static constructors.

[zack-alex]

Right, the example was a bit confusing. Here's a new one with an explicit entry point:

struct ClassA {
  char data_a[400000]; // If this is commented, we get "42" as expected.
};

struct ClassB {
  ClassB() {
    value = 33;
  }
  char data_b[400000]; // If this is commented, we get "RuntimeError: out of bounds memory access".
  int value;
};

ClassB global{};

__attribute__((export_name("example_func"))) int example_func() {
  global.value += 9;
  ClassA local{};
  return global.value; // Should be "42", but we get "0".
}

Command log:

+ clang-17 --version
clang version 17.0.6
Target: aarch64-apple-darwin23.6.0
Thread model: posix
InstalledDir: /nix/store/glzjy0xq2xhmwg6pc8c1pdk7gln127wj-clang-17.0.6/bin
+ clang-17 -std=c++20 -pedantic -Wall -Wno-unused-variable --target=wasm32 -Wl,--no-entry --no-standard-libraries -mbulk-memory -O0 -o main.wasm main.cpp
+ wasmer --entrypoint example_func main.wasm
0

Note that static initialisation is, in fact, performed; its invocation is prepended to example_func.

Here's the generated Wasm

(module
  (type $t0 (func))
  (type $t1 (func (param i32) (result i32)))
  (type $t2 (func (result i32)))
  (func $__wasm_call_ctors (type $t0)
    (call $_GLOBAL__sub_I_main.cpp))
  (func $__cxx_global_var_init (type $t0)
    (local $l0 i32)
    (local.set $l0
      (i32.const 1024))
    (drop
      (call $ClassB::ClassB__
        (local.get $l0)))
    (return))
  (func $ClassB::ClassB__ (type $t1) (param $p0 i32) (result i32)
    (local $l1 i32) (local $l2 i32) (local $l3 i32) (local $l4 i32) (local $l5 i32)
    (local.set $l1
      (global.get $__stack_pointer))
    (local.set $l2
      (i32.const 16))
    (local.set $l3
      (i32.sub
        (local.get $l1)
        (local.get $l2)))
    (i32.store offset=12
      (local.get $l3)
      (local.get $p0))
    (local.set $l4
      (i32.load offset=12
        (local.get $l3)))
    (local.set $l5
      (i32.const 33))
    (i32.store offset=400000
      (local.get $l4)
      (local.get $l5))
    (return
      (local.get $l4)))
  (func $example_func__ (type $t2) (result i32)
    (local $l0 i32) (local $l1 i32) (local $l2 i32) (local $l3 i32) (local $l4 i32) (local $l5 i32) (local $l6 i32) (local $l7 i32) (local $l8 i32) (local $l9 i32) (local $l10 i32) (local $l11 i32) (local $l12 i32) (local $l13 i32)
    (local.set $l0
      (global.get $__stack_pointer))
    (local.set $l1
      (i32.const 400000))
    (local.set $l2
      (i32.sub
        (local.get $l0)
        (local.get $l1)))
    (global.set $__stack_pointer
      (local.get $l2))
    (local.set $l3
      (i32.const 0))
    (local.set $l4
      (i32.load offset=401024
        (local.get $l3)))
    (local.set $l5
      (i32.const 9))
    (local.set $l6
      (i32.add
        (local.get $l4)
        (local.get $l5)))
    (local.set $l7
      (i32.const 0))
    (i32.store offset=401024
      (local.get $l7)
      (local.get $l6))
    (local.set $l8
      (i32.const 400000))
    (local.set $l9
      (i32.const 0))
    (memory.fill
      (local.get $l2)
      (local.get $l9)
      (local.get $l8))
    (local.set $l10
      (i32.const 0))
    (local.set $l11
      (i32.load offset=401024
        (local.get $l10)))
    (local.set $l12
      (i32.const 400000))
    (local.set $l13
      (i32.add
        (local.get $l2)
        (local.get $l12)))
    (global.set $__stack_pointer
      (local.get $l13))
    (return
      (local.get $l11)))
  (func $_GLOBAL__sub_I_main.cpp (type $t0)
    (call $__cxx_global_var_init)
    (return))
  (func $example_func____.command_export_ (export "example_func") (type $t2) (result i32)
    (call $__wasm_call_ctors)
    (call $example_func__))
  (table $T0 1 1 funcref)
  (memory $memory (export "memory") 8)
  (global $__stack_pointer (mut i32) (i32.const 466576)))

[macsencasaus]

Looks like a stack overflow, you can increase the stack size with the compile flag: -Wl,-z,stack-size=524288.
This returns 42.

[zack-alex]

Okay, in such case it's not a miscompilation but I have several UX suggestions:

• How do I even find out the default stack size? wasm32-unknown-wasi-ld --help doesn't mention stack-size option at all. Neither does the web manual.
• So the default stack is less than 400KB? So little? Any plans to increase it?
• Any plans to do something so that stack overflow causes a more guaranteed and clear error (like out-of-memory access) instead of unpredictable corruption? I'm not an expert here, so pardon me if my suggestions are stupid: maybe change the memory layout so that the stack is the first thing in memory and it grows downwards, so that stack overflow leads to out-of-memory access? Or maybe use wasm multi-memory extension and move the stack to a separate memory?

[macsencasaus]

The stack-size option is a linker option extension. That is why it is prefixed with -z which I think are mostly inherited from lld. You can see the stack-size option in the man pages of lld with man lld. I'm unsure about the rest of your questions.

[sbc100]

• Any plans to do something so that stack overflow causes a more guaranteed and clear error (like out-of-memory access) instead of unpredictable corruption? I'm not an expert here, so pardon me if my suggestions are stupid: maybe change the memory layout so that the stack is the first thing in memory and it grows downwards, so that stack overflow leads to out-of-memory access? Or maybe use wasm multi-memory extension and move the stack to a separate memory?

The stack size defaults if 64KiB (which is 1 page of wasm memory). It don't know of any plans to change the default since wasm is orten/normally used in resource constrained environments. However its easy to override with -zstack-size=xx as has already be pointed out.

There are also debugging aids in emscripten such as emscriptens -sSTACK_OVERFLOW_CHECK: https://emscripten.org/docs/tools_reference/settings_reference.html#stack-overflow-check. This in turn relies on a binaryen pass which can also be used outside of emscripten: https://github.com/WebAssembly/binaryen/blob/729ea41d145d369b203dca6f70b251ea365cb3d0/src/passes/StackCheck.cpp#L17-L21

[sbc100]

Closing for now. Feel free to re-open if you disagree (or open another issue if you want to propose a new default value).