wasp-opt makes bcrypt code slower

I'm not sure if this is issues you want to receive, but I ran into this and thought that it could be good to report. Feel free to close this if this is not relevant.

I'm compiling Openwall's BCrypt implementation into WASM with clang. I'm compiling with -O3 and -flto. Right out from clang I get the following performance:

wasm Openwall x 4.11 ops/sec ±0.13% (15 runs sampled)

When I then run wasm-opt --enable-bulk-memory -O3 on the resulting wasm file, the performance decreases:

wasm Openwall2 x 3.73 ops/sec ±0.06% (14 runs sampled)

I get the same results when passing -O4 to wasm-opt.

You can find my project here:

https://github.com/LinusU/cwasm-openwall-bcrypt

There is a Dockerfile that builds the entire project, so it's easy to experiment with passing different flags to clang. I'm using the lines below to add and run wasm-opt. Do note that these needs to be added after the clang invocation, since clang will automatically invoke wasm-opt if it's present in the path.

RUN curl -L https://github.com/WebAssembly/binaryen/releases/download/version_110/binaryen-version_110-x86_64-linux.tar.gz | tar xzk --strip-components=1 -C /
RUN wasm-opt --enable-bulk-memory -O3 bcrypt.wasm -o bcrypt.wasm

Start by reproducing the Dockerfile workflow with the reported wasm-opt --enable-bulk-memory -O3 command and compare its performance with the clang-produced wasm. Use the reported bcrypt.wasm input and optimization flags to isolate the slowdown. Done means identifying and addressing the optimization regression, with performance results showing the optimized output is not slower.