Large Binary Instruments Profile Data

[milend]

I've recorded profiling data (zld.trace.zip) for a large binary which takes several minutes to link. In terms of numbers:

• ld64
  • 225.64s user 17.44s system 94% cpu 4:17.81 total
  • 219.63s user 17.33s system 93% cpu 4:14.00 total
  • 221.44s user 21.93s system 88% cpu 4:35.26 total
• zld
  • 213.75s user 14.14s system 118% cpu 3:12.50 total
  • 218.63s user 16.22s system 115% cpu 3:22.62 total
  • 213.71s user 14.74s system 118% cpu 3:11.99 total

zld is around 1min faster (~23% faster). Machine was rebooted before taking the measurements and the samples taken with hot file caches (i.e., linkers ran before recording the samples above). Machine is a 15" MacBook Pro 2017, i7 3.1GHz, 16GB RAM and 1TB SSD.

[michaeleisel]

Thank you!

[milend]

@michaeleisel: here's another trace of linking the same exact binary without passing an order file. This cuts the linking time from around 3m15s to 56s (outside Instruments). There's still a large section which is single-threaded, so I'm wondering if it can be parallelised – it would result in a huge perf win.

zld-no-order.trace.zip

[tcpthecode]

@milend how did you pass order file into linker ? In my case, the binary layout is just the same after defined order file for linking 🤔

[milend]

@tcpthecode We use the -order_file parameter.

[tcpthecode]

Thanks @michaeleisel 👍