WASM performance is ~45% slower compared to 2.9

[tzimnoch]

Sample performance tests can be checked out from https://github.com/tzimnoch/dds/tree/performance_demo

The 2.9 WASM files were build from https://github.com/dds-bridge/dds.git using the following command.

$ em++ --version
emcc (Emscripten gcc/clang-like replacement + linker emulating GNU ld) 5.0.3 (285c424dfa9e83b03cf8490c65ceadb7c45f28eb)
Copyright (C) 2026 the Emscripten authors (see AUTHORS.txt)
[...]

em++ -O2 \
  -s WASM=1 \
  -s EXPORTED_FUNCTIONS='["_malloc","_free","_SetMaxThreads","_SolveBoard","_SolveBoardPBN","_CalcDDtable","_CalcDDtablePBN"]' \
  -s EXPORTED_RUNTIME_METHODS='["ccall","cwrap","setValue","getValue"]' \
  -s INITIAL_MEMORY=67108864 \
  -s MODULARIZE=1 \
  -s EXPORT_NAME="DDS" \
  -s USE_PTHREADS=0 \
  -D DDS_THREADS=1 \
  -I src/ \
  src/*.cpp \
  -o dds.js

Using Firefox on Fedora, solving the same set of 100 random boards took:
New solver: 492925.0ms (4929.25ms avg). Legacy: 263241.0ms (2632.41ms avg).

Brave on Fedora had a similar difference in relative performance, though significantly better than Firefox.
New solver: 124909.7ms (1249.10ms avg). Legacy: 70628.5ms (706.28ms avg).

As of yet I've not been able to account for the difference in performance nor have I been successful in tweaking the compile/linker options in this project to achieve similar performance as seen in prior versions.

I have not verified whether a similar difference in performance exists in the native libraries, but I suspect the greatest performance gain will come from changing the way the WASM artifact is built or used.

[tameware]

C++ performance is similarly slower, even for the single-threaded case. This seems due to an architecture change that I do not pretend to understand.