minimodem binary works in wasmtime, doesn't work in wasmer

[jcaesar]

Describe the bug

This wasm-wasi binary (compiled from here) works fine on wasmtime:

echo This message will get through | wasmtime run -- minimodem.wasm --tx --tx-carrier --stdio --float-samples same | wasmtime run -- minimodem.wasm --rx --tx-carrier --stdio --float-samples same
### CARRIER 521 @ 2083.3 Hz ###
This message will get through

### NOCARRIER ndata=46 confidence=141751.797 ampl=45.977 bps=521.03 (0.0% fast) ###

But bugs on wasmer

echo This message wont get through | wasmer run -- minimodem.wasm --tx --tx-carrier --stdio --float-samples same | wasmer run -- minimodem.wasm --rx --tx-carrier --stdio --float-samples same
### CARRIER 521 @ 2083.3 Hz ###
R�iln w�4�)
### NOCARRIER ndata=36 confidence=90307.961 ampl=41.893 bps=515.86 (1.0% slow) ###

Steps to reproduce

1. Download binary linked above
2. Run commands shown above

Expected behavior

Shouldn't fail to decode audio

Actual behavior

Fails to decode audio

Additional context

• This might not be a bug in wasmer and just minimodem doing something awful in C
• I inserted a debug print into the code that reads samples, which showed that at some point, the sample values that wasmer and wasmtime read are different. So this might have a similar cause as wasi fd_read sets bogous n_bytes_read value / ignores buffer length on webassembly.sh webassembly.sh#112, but that bug only appears on webassembly.sh
• This is probably a pain to debug. :(

[ptitSeb]

I reproduced the issue, even with current wasmer (almost 3.0) code.

I built a native version of minimodem also, and found that only the "rx" part is broken. The "tx" part seems to be working as intended.
Also, I tried with all the compiler and the issue is still there, so probably something on the WASI side.

I'll continue my investigations.

[ptitSeb]

Ok, I found the issue. It was the fd_read WASI syscall. When reading on multiple iovs, if a read is partial (so not all the wanted bytes are read), the function would still continue to try reading on subsequent iov (eventualy reading more bytes but not in the right place).

[jcaesar]

6d08c88 does indeed fix it. Ty.

[jcaesar]

This exact same issue reappeared (though the cause is probably different, last time the --rx side was broken, this time it's the --tx side [Edit:] On master, it's rx that's broken again. Ookay?). Please do tell if you'd prefer I open a new issue.

I'll try to bisect it if I find the time. Should I also add a minimodem test case? It seems to be good at finding bugs.

[ptitSeb]

Ok, I tested with current master, and indeed it's borken.
I also tested "tx" and "rx" part separatly, and they seems both broken. If I redirect the "tx" part to a file, that file is different from a reference file generated wth wasmer 2.3 for example.

I'll try to digg in and see what is happening.

[ptitSeb]

The second part of the fix has been merged with #3741

[jcaesar]

Wow, fast.

[ptitSeb]

And it will be tested now, it's been added to snapshot tests. No more regressions :)