WebAssembly doesn't have 128-bit integers, so on wasm targets LLVM implements 128-bit multiplies with a libcall to __multi3. That's slower than just doing arithmetic on smaller units.
I patched the rsa crate to remove the u64_digit feature from the num-bigint-dig dependency, and tested the patched version with the blind-rsa-signatures crate. The toy demo app I tested ran 15% faster under Wasmtime with this patch.
It would be nice if either this crate or num-bigint-dig would automatically determine what limb size to use based on the build target. Barring that, perhaps this crate should expose its own u64_digit feature that forwards to num-bigint-dig.
WebAssembly doesn't have 128-bit integers, so on wasm targets LLVM implements 128-bit multiplies with a libcall to
__multi3. That's slower than just doing arithmetic on smaller units.I patched the
rsacrate to remove theu64_digitfeature from thenum-bigint-digdependency, and tested the patched version with theblind-rsa-signaturescrate. The toy demo app I tested ran 15% faster under Wasmtime with this patch.It would be nice if either this crate or
num-bigint-digwould automatically determine what limb size to use based on the build target. Barring that, perhaps this crate should expose its ownu64_digitfeature that forwards tonum-bigint-dig.