Closed koute closed 2 months ago
Using varints instead of u32s for the instruction stream costs us ~15% of compilation performance, which is relatively huge. We should switch to a SIMD-friendly format to make the decoding even faster than just using u32.
u32
An approach inspired by something like this would probably work pretty well: https://arxiv.org/abs/1709.08990
Using varints instead of
u32
s for the instruction stream costs us ~15% of compilation performance, which is relatively huge. We should switch to a SIMD-friendly format to make the decoding even faster than just usingu32
.An approach inspired by something like this would probably work pretty well: https://arxiv.org/abs/1709.08990