[Old EH] JS stack unwinding support corrupts some floats

[hoodmane]

main.c++

#include <stdexcept>
#include <stdio.h>
using namespace std;
typedef unsigned int uint;
uint float_to_uint(float f);
float uint_to_float(uint i);

int main() {
    uint x = 4287976789;
    float f;
    try {
        f = uint_to_float(x);
    } catch(runtime_error& e) {
        printf("Caught runtime error...");
    }
    uint y = float_to_uint(f);
    printf("x: %ud, y: %ud, y - x: %ud\n", x, y, y - x);
    if (x != y) {
      printf("Oops!\n");
      return 1;
    }
    return 0;
}

helper.c++

typedef unsigned int uint;

typedef union { 
    uint i; 
    float f; 
} U; 

uint float_to_uint(float f) {
    U u;
    u.f = f;
    return u.i;
}

float uint_to_float(uint i) {
    U u;
    u.i = i;
    return u.f;
}

Running it

Compile and run with

em++ main.c++ -fexceptions -c && em++ helper.c++ -fexceptions -c &&  em++ main.o helper.o -fexceptions -o repro.js && node repro.js

it prints:

x: 4287976789d, y: 4292171093d, y - x: 4194304d
Oops!

On the other hand, if we compile and run with -fwasm-exceptions:

em++ main.c++ -fwasm-exceptions -c && em++ helper.c++ -fwasm-exceptions -c &&  em++ main.o helper.o -fwasm-exceptions -o repro.js && node repro.js

It prints the expected output:

x: 4287976789d, y: 4287976789d, y - x: 0d

Emscripten version

$ emcc -v
emcc (Emscripten gcc/clang-like replacement + linker emulating GNU ld) 3.1.68 (ceee49d2ecdab36a3feb85a684f8e5a453dde910)
clang version 20.0.0git (https:/github.com/llvm/llvm-project 5cc64bf60bc04b9315de3c679eb753de4d554a8a)
Target: wasm32-unknown-emscripten
Thread model: posix

[kripken]

Hmm, the integer here is turned into a NaN by uint_to_float. And then the difference between old JS EH and new wasm EH is that the former goes through JS. Avoiding the try-catch around that call leads to the same behavior as new wasm EH, as nothing is thrown, so this is a bug in old EH (and not related to new EH).

The NaN gets canonicalized at the JS boundary, I believe (calling invoke_fi returns that NaN as a JS result) which is what leads to this issue. That is, we need a very particular pattern of NaN bits here, not the canonical NaN that JS (but not wasm) forces everything into.

We could in theory fix this by making invoke_fi return an integer and not a float directly, and we could do a reinterpret on the wasm side. That is, we could avoid sending floats through JS. I don't think that would be too much work, but OTOH this is for old EH, which is being phased out anyhow, so perhaps it isn't worth it?

[hoodmane]

Well anyone who uses rust is stuck with the old error handling until someone updates the rust try intrinsic to emit ir compatible with the new error handling:
#19612

[kripken]

@hoodmane is that an issue on the Rust side, or LLVM? (the link is to an Emscripten issue, but I'm not sure I understood from it where the underlying problem is)

In general, getting Rust to emit IR compatible with new wasm EH sounds very important for any Rust/C++ combination, Emscripten or otherwise. Is there a Rust issue for that?

With all that said, I wouldn't be opposed to fixing float invokes for old EH, using the approach I mentioned above, if someone has time for it.

[hoodmane]

getting Rust to emit IR compatible with new wasm EH sounds very important

Yeah I would love to switch Pyodide to use the new wasm EH and this is the only blocker.

The issue is in the rust compiler. Specifically, the try intrinsic needs to be updated:
https://github.com/rust-lang/rust/blob/master/compiler/rustc_codegen_llvm/src/intrinsic.rs?plain=1#L971

[kleisauke]

The tracking issue for WebAssembly EH in Rust is at rust-lang/rust#118168. Looking at PR rust-lang/rust#111322, I'm unsure why it was conditionally excluded for Emscripten:
https://github.com/rust-lang/rust/blob/0037048da8ed19feef7ee1437f123a88a7342b33/compiler/rustc_codegen_ssa/src/base.rs#L389-L391

Could you try to remove the && sess.target.os != "emscripten" there? This would make it use codegen_wasm_try instead of codegen_emcc_try.
https://github.com/rust-lang/rust/blob/0037048da8ed19feef7ee1437f123a88a7342b33/compiler/rustc_codegen_llvm/src/intrinsic.rs#L666-L669

[kripken]

Thanks @kleisauke !

@walkingeyerobot fyi, the last comment has some details that might be relevant for future Rust/Emscripten work.

[hoodmane]

Well using that is good progress. I can get it to link with the new codegen and the try intrinsic looks perfect, but there are problems elsewhere in rust/library/panic_unwind. First I have to change the argument to __cxa_begin_catch I think it's supposed to be the wasm exception argument, the rust code for old EH dereferences it once first. This hits all the right code paths but then I'm running into the problem that exception_header->adjustedPtr is null. I'm not sure who was supposed to set adjustedPtr, I'll have to look into it more.

[hoodmane]

Looks like adjustedPtr was supposed to be filled in by _Unwind_CallPersonality, but as far as I can tell it never gets called. In ir, the catchpad looks like:

catchpad:
   %tok = catchpad within %cs [null]
   %ptr = call @llvm.wasm.get.exception(token %tok)
   %sel = call @llvm.wasm.get.ehselector(token %tok)
   call %catch_func(%data, %ptr)
   catchret from %tok to label %caught

Was @llvm.wasm.get.exception supposed to call it? The catch block in the generated wasm only has:

catch $tag0
      local.set $var4
      local.get $var3
      global.set $__stack_pointer
      local.get $var1
      local.get $var4
      local.get $var2
      call_indirect (param i32 i32)
      i32.const 1
      local.set $var5
      local.get $var5
end

and the call_indirect is call $catch_func.

[hoodmane]

Indeed WasmEHPrepare is supposed to turn @llvm.wasm.get.exception into a call to _Unwind_CallPersonality:

WebAssembly exception handling uses Windows exception IR for the middle level
representation. This pass does the following transformation for every
catchpad block:
(In C-style pseudocode)

• Before:

 catchpad ...
 exn = wasm.get.exception();
 selector = wasm.get.selector();
 ...

• After:

  catchpad ...
  exn = wasm.catch(WebAssembly::CPP_EXCEPTION);
  // Only add below in case it's not a single catch (...)
  wasm.landingpad.index(index);
  __wasm_lpad_context.lpad_index = index;
  __wasm_lpad_context.lsda = wasm.lsda();
  _Unwind_CallPersonality(exn);
  selector = __wasm_lpad_context.selector;
  ...

https://github.com/llvm/llvm-project/blob/main/llvm/lib/CodeGen/WasmEHPrepare.cpp?plain=1#L13-L33

I wonder if I can get llvm to print the result of this pass...

[hoodmane]

After the WasmEHPrepare pass the ir looks like:

catchpad:                                         ; preds = %catchswitch
  %catchpad2 = catchpad within %catchswitch1 [ptr null]
  %exn = call ptr @llvm.wasm.catch(i32 0)
  call void %2(ptr %1, ptr %exn) [ "funclet"(token %catchpad2) ]
  catchret from %catchpad2 to label %caught

It seems we took that !NeedPersonality branch in the pass:
https://github.com/llvm/llvm-project/blob/main/llvm/lib/CodeGen/WasmEHPrepare.cpp?plain=1#L342-L349

  // In case it is a catchpad with single catch (...) or a cleanuppad, we don't
  // need to call personality function because we don't need a selector.
  if (!NeedPersonality) {
    if (GetSelectorCI) {
      assert(GetSelectorCI->use_empty() &&
             "wasm.get.ehselector() still has uses!");
      GetSelectorCI->eraseFromParent();
    }
    return;
  }

So as the comment says, this would normally happen in C++ when there is a blanket catch (...) and so the return value of __cxa_begin_catch is unused. Rust only has one llvm catch block to handle all the different cases, but it wants to know what exception was thrown. I think the simplest solution for this is for Rust to explicitly call the personality function in this case, at least if that doesn't make the linker angry...

[hoodmane]

Adding the call to _Unwind_CallPersonality doesn't help because _Unwind_GetLanguageSpecificData returns null.