Add compiling for specific architectures section to README

[sayrer]

This did get a win on the matcher--Go's AMD default is really old stuff.

[timbray]

Might it add value to say what you tested it on and what level you used and what the observed perf boost was, just as an example to make this concrete?

[timbray]

… by which I mean "I'm curious about what you tested on and what you saw" and I suspect if I am lots of other people will be too.

[sayrer]

 Final result: v1 vs v3, -count=6, benchstat

  Overall geomean: −0.75% (v3 faster). The instruction-set bump produces a small but real, 
  consistent improvement — almost everything that moved, moved down, with most changes
  in the 1–5% range.

  Notable significant wins (p < 0.05, n=6)

  ┌─────────────────────────────────────────────┬──────────┬──────────┬────────────────┐
  │                  Benchmark                  │    v1    │    v3    │       Δ        │              
  ├─────────────────────────────────────────────┼──────────┼──────────┼────────────────┤
  │ Workload_ParallelMatchers/G=8               │ 1075 ns  │ 980 ns   │ −8.9%          │
  ├─────────────────────────────────────────────┼──────────┼──────────┼────────────────┤
  │ Nfa2Dfa/two_stars                           │ 270.3 µs │ 253.0 µs │ −6.4%          │
  ├─────────────────────────────────────────────┼──────────┼──────────┼────────────────┤
  │ Workload_ExactString                        │ 300 ns   │ 285 ns   │ −5.2%          │
  ├─────────────────────────────────────────────┼──────────┼──────────┼────────────────┤
  │ Workload_LiteralInRegex                     │ 117 ns   │ 112 ns   │ −4.5%          │
  ├─────────────────────────────────────────────┼──────────┼──────────┼────────────────┤
  │ Nfa2Dfa/five_stars                          │ 87.8 µs  │ 84.1 µs  │ −4.2%          │
  ├─────────────────────────────────────────────┼──────────┼──────────┼────────────────┤
  │ Workload_ParallelMatchers/G=32              │ 1066 ns  │ 1020 ns  │ −4.3%          │
  ├─────────────────────────────────────────────┼──────────┼──────────┼────────────────┤
  │ ShellstyleSimpleWildcardScaling (all sizes) │ ~600 ns  │ ~580 ns  │ −1.6% to −3.2% │
  ├─────────────────────────────────────────────┼──────────┼──────────┼────────────────┤
  │ Workload_SingleShellstyle                   │ 663 ns   │ 640 ns   │ −3.5%          │
  ├─────────────────────────────────────────────┼──────────┼──────────┼────────────────┤
  │ CityLots                                    │ 10.70 µs │ 10.56 µs │ −1.3%          │
  └─────────────────────────────────────────────┴──────────┴──────────┴────────────────┘

  Regressions (small, likely noise)
  
  ┌────────────────────────────────────────────┬───────┬─────────────────────────────────────────────────────────────┐
  │                 Benchmark                  │   Δ   │                            note                             │
  ├────────────────────────────────────────────┼───────┼─────────────────────────────────────────────────────────────┤
  │ Workload_ManyOverlappingWildcards/N=64     │ +5.9% │ high variance (±6%), inconsistent with N=8/32 which         │
  │                                            │       │ improved                                                    │
  ├────────────────────────────────────────────┼───────┼─────────────────────────────────────────────────────────────┤
  │ ShellstyleWidePatternsScaling/patterns=512 │ +2.0% │ ±9% at v1 — noisy                                           │
  ├────────────────────────────────────────────┼───────┼─────────────────────────────────────────────────────────────┤
  │ ShellstyleZWJEmoji/patterns=8              │ +0.8% │ within noise                                                │
  ├────────────────────────────────────────────┼───────┼─────────────────────────────────────────────────────────────┤
  │ _JsonFlattener_MiddleNestedField           │ +0.9% │ flattener is unaffected by ISA, as expected                 │
  └────────────────────────────────────────────┴───────┴─────────────────────────────────────────────────────────────┘

  The JSON-flattener benchmarks are essentially flat (±1%, mostly "~"), which makes sense — 
  that path is byte-scanning with no arithmetic the compiler would vectorize or BMI2-optimize.

  Bottom line

  v4 cannot be measured on this box — the Threadripper 2920X has no AVX‑512, and Go refuses 
  to run a v4 binary. You'd need a Zen 4+ or Intel Skylake-X/Ice-Lake+ machine to test it.

AMD Threadripper 2920X. Got the idea from Daniel Lemire, and had Claude do the benchmarking duties while I did other things.

[sayrer]

I had other runs that got like 9.6% on Workload_ParallelMatchers/G=8, but this is not a pristine quiet benchmark machine.