Low Level Tools/2019Q2

Nick

rustc pipelining: Implement "pipelined" rustc compilation (compiler side).

• 100% completed.
• Don't generate unnecessary rmeta files. A precursor.
• In JSON output, emit a directive after metadata is generated. The part that lets rustc tell Cargo when metadata has been written.
• Emit metadata files earlier. So that significant crate overlapping can occur.
• Tracking issue for stabilizing "pipelined compilation" has been opened.
• Announcement email to dev-platform
• Speed-ups depend on program structure and configuration, but up to 1.84x has been observed.
• Speedups for opt builds of notable projects on my 14-core Linux box:
  • webrender: 1.27x
  • cranelift-codegen: 1.20x
  • Firefox: 1.07x
  • script-servo: 1.07x
  • style-servo: 1.06x

Ship GWP.

Land 5 changes that speed up or reduce peak memory usage of rustc.

• Increase `Span` from 4 bytes to 8 bytes. A big win across lots of workloads, especially larger crates: some script-servo builds are 15% faster.
• Avoid `hygiene_data` lookups. Up to 10% win on packed-simd, up to 3% on numerous others.
• Reduce the `DepNode` pre-allocation ratio. Reduces peak memory use by 20 MiB on some workloads.
• Use `Symbol` more. Up to 1% speed-up, as well as making the code nicer. Should also help with parallel rustc.
• Use `Symbol` even more. Up to 1% speed-up. Should also help with parallel rustc.