free or reuse unconditionally when value is unique

[folkertdev]

No description provided.

[folkertdev]

the benchmarks fail on deriv, but locally I get these results

hyperfine ./old ./new
Benchmark #1: ./old
  Time (mean ± σ):     397.3 ms ±  11.1 ms    [User: 362.5 ms, System: 33.6 ms]
  Range (min … max):   387.1 ms … 415.5 ms    10 runs
 
Benchmark #2: ./new
  Time (mean ± σ):     372.8 ms ±   2.6 ms    [User: 342.1 ms, System: 30.0 ms]
  Range (min … max):   369.2 ms … 377.5 ms    10 runs
 
Summary
  './new' ran
    1.07 ± 0.03 times faster than './old'

so I'm inclined to ignore the CI failure

[folkertdev]

results for cfold, with n = 20 (using a c platform to work around zig's limitations on recursion)

Benchmark #1: ./old
  Time (mean ± σ):     867.7 ms ±  20.2 ms    [User: 480.9 ms, System: 386.8 ms]
  Range (min … max):   854.1 ms … 921.5 ms    10 runs
  
Benchmark #2: ./new
  Time (mean ± σ):     752.6 ms ±  21.0 ms    [User: 447.1 ms, System: 305.3 ms]
  Range (min … max):   736.4 ms … 802.6 ms    10 runs
 
Summary
  './new' ran
    1.15 ± 0.04 times faster than './old'

and deriv with n = 10

Benchmark #1: ./old
  Time (mean ± σ):      3.018 s ±  0.018 s    [User: 2.781 s, System: 0.236 s]
  Range (min … max):    2.993 s …  3.051 s    10 runs
 
Benchmark #2: ./new
  Time (mean ± σ):      2.939 s ±  0.022 s    [User: 2.688 s, System: 0.247 s]
  Range (min … max):    2.918 s …  2.976 s    10 runs
 
Summary
  './new' ran
    1.03 ± 0.01 times faster than './old'

so to me that is sufficient proof that there is no regression for deriv, and we actually get a nice speedup for cfold.

[JTeeuwissen]

In addition, the IR does not have any irregular changes.