bd870a9d6c
We weren't using the "union" cargo feature for the smallvec crate, which reduces the size of a SmallVec by one machine word. This feature requires Rust 1.49 but we already require much newer versions. When using Wasmtime to compile pulldown-cmark from Sightglass, this saves a decent amount of memory allocations and writes. According to `valgrind --tool=dhat`: - 6.2MiB (3.69%) less memory allocated over the program's lifetime - 0.5MiB (4.13%) less memory allocated at maximum heap size - 5.5MiB (1.88%) fewer bytes written to - 0.44% fewer instructions executed Sightglass reports a statistically significant runtime improvement too: compilation :: cycles :: benchmarks/pulldown-cmark/benchmark.wasm Δ = 24379323.60 ± 20051394.04 (confidence = 99%) shrink-abiarg-0406da67c.so is 1.01x to 1.13x faster than main-be690a468.so! [227506364 355007998.78 423280514] main-be690a468.so [227686018 330628675.18 406025344] shrink-abiarg-0406da67c.so compilation :: cycles :: benchmarks/spidermonkey/benchmark.wasm Δ = 360151622.56 ± 278294316.90 (confidence = 99%) shrink-abiarg-0406da67c.so is 1.01x to 1.07x faster than main-be690a468.so! [8709162212 8911001926.44 9535111576] main-be690a468.so [5058015392 8550850303.88 9282148438] shrink-abiarg-0406da67c.so compilation :: cycles :: benchmarks/bz2/benchmark.wasm Δ = 6936570.28 ± 6897696.38 (confidence = 99%) shrink-abiarg-0406da67c.so is 1.00x to 1.08x faster than main-be690a468.so! [155810934 175260571.20 234737344] main-be690a468.so [119128240 168324000.92 257451074] shrink-abiarg-0406da67c.so
This crate provides a straightforward way to create a
Cranelift IR function and fill it with
instructions translated from another language. It contains an SSA construction
module that provides convenient methods for translating non-SSA variables into
SSA Cranelift IR values via use_var and def_var calls.