Add frexpf128, ilogbf16, and ilogbf128

[tgross35]

ci: allow-regressions

[tgross35]

Current regression on aarch64 shows

  icount::icount_bench_frexp_group::icount_bench_frexp logspace:
    Callgrind: Instructions (7719 -> 8219): +6.47752% exceeds limit of +5.00000%
  icount::icount_bench_frexpf_group::icount_bench_frexpf logspace:
    Callgrind: Instructions (7751 -> 8251): +6.45078% exceeds limit of +5.00000%

Previously there was a single function, but with the generic there is:

$ cargo asm -p libm --lib --target aarch64-unknown-linux-gnu frexp 0
   Compiling libm v0.2.16 (~/projects/compiler-builtins/libm)
    Finished `release` profile [optimized] target(s) in 1.11s

.section .text.libm::math::frexp::frexp,"ax",@progbits
        .globl  libm::math::frexp::frexp
        .p2align        2
.type   libm::math::frexp::frexp,@function
libm::math::frexp::frexp:
        .cfi_startproc
        b libm::math::generic::frexp::frexp::<f64>

Not sure why it isn't getting inlined, even #[inline(always)] doesn't help.

[tgross35]

Oh, duh, it won't inline because the generic frexp calls itself. x86 has the regression too, just not over the threshold:

icount::icount_bench_frexp_group::icount_bench_frexp logspace:setup_frexp()
  Baselines:                      softfloat|softfloat (old)
  Instructions:                       11689|11189                (+4.46867%) [+1.04469x]
  L1 Hits:                            13274|12777                (+3.88980%) [+1.03890x]
  LL Hits:                                3|1                    (+200.000%) [+3.00000x]
  RAM Hits:                               6|5                    (+20.0000%) [+1.20000x]
  Total read+write:                   13283|12783                (+3.91144%) [+1.03911x]
  Estimated Cycles:                   13499|12957                (+4.18307%) [+1.04183x]
icount::icount_bench_frexpf_group::icount_bench_frexpf logspace:setup_frexpf()
  Baselines:                      softfloat|softfloat (old)
  Instructions:                       10726|10226                (+4.88950%) [+1.04889x]
  L1 Hits:                            12321|11824                (+4.20332%) [+1.04203x]
  LL Hits:                                2|1                    (+100.000%) [+2.00000x]
  RAM Hits:                               5|5                    (No change)
  Total read+write:                   12328|11830                (+4.20964%) [+1.04210x]
  Estimated Cycles:                   12506|12004                (+4.18194%) [+1.04182x]

I'm just going to accept the regression here since I'll be following up with a change to the algorithm to support f16.

[tgross35]

#1092