Intermittent test failure in math-effect test

[Keno]

Failure is test order dependent and reproducible with

JULIA_CPU_THREADS=1 ../julia runtests.jl LinearAlgebra/svd LinearAlgebra/pinv LinearAlgebra/givens LinearAlgebra/structuredbroadcast numbers copy math
Running parallel tests with:
  nworkers() = 1
  nthreads() = 1
  Sys.CPU_THREADS = 1
  Sys.total_memory() = 1007.822 GiB
  Sys.free_memory() = 992.748 GiB

Test                          (Worker) | Time (s) | GC (s) | GC % | Alloc (MB) | RSS (MB)
LinearAlgebra/svd                  (1) |        started at 2022-06-22T03:20:39.381
LinearAlgebra/svd                  (1) |    50.03 |   2.66 |  5.3 |    7520.24 |   665.83
LinearAlgebra/pinv                 (1) |        started at 2022-06-22T03:21:29.526
LinearAlgebra/pinv                 (1) |     9.04 |   0.66 |  7.3 |    1456.71 |   839.06
LinearAlgebra/givens               (1) |        started at 2022-06-22T03:21:38.564
LinearAlgebra/givens               (1) |     4.25 |   0.21 |  5.0 |     441.16 |   848.18
LinearAlgebra/structuredbroadcast  (1) |        started at 2022-06-22T03:21:42.814
LinearAlgebra/structuredbroadcast  (1) |    33.33 |   2.27 |  6.8 |    4345.93 |   909.84
numbers                            (1) |        started at 2022-06-22T03:22:16.142
numbers                            (1) |    48.12 |   1.14 |  2.4 |    4384.62 |  1356.91
copy                               (1) |        started at 2022-06-22T03:23:04.259
copy                               (1) |     2.52 |   0.09 |  3.4 |     277.91 |  1356.91
math                               (1) |        started at 2022-06-22T03:23:06.782
┌ Warning: bad effects found for sin(::Float32)
│   eff = (+c,+e,?n,?t,+s)
└ @ Main.Test6Main_math ~/julia/test/math.jl:1467
┌ Warning: bad effects found for cos(::Float32)
│   eff = (+c,+e,?n,?t,+s)
└ @ Main.Test6Main_math ~/julia/test/math.jl:1467
┌ Warning: bad effects found for tan(::Float32)
│   eff = (+c,+e,?n,?t,+s)
└ @ Main.Test6Main_math ~/julia/test/math.jl:1467
math                               (1) |    36.14 |   1.31 |  3.6 |    4391.28 |  1625.89

Test Summary: |    Pass  Broken    Total     Time
  Overall     | 3493226       3  3493229  3m04.5s
    SUCCESS

Originally posted by @Keno in #45670 (comment)

[staticfloat]

@aviatesk and I managed to squeeze this down to reproducible with just:

Base.runtests(["numbers", "math"]; ncores=1)

[martinholters]

Looks like sign(pi) (contained in numbers test) is the offender.
Without:

julia> eff = Base.infer_effects(sin, (Float32,))
(+c,+e,?n,+t,+s)

julia> Core.Compiler.is_foldable(eff)
true

Fresh session:

julia> sign(pi)
1.0

julia> eff = Base.infer_effects(sin, (Float32,))
(+c,+e,?n,?t,+s)

julia> Core.Compiler.is_foldable(eff)
false

[martinholters]

Drilling a bit into sign(pi), I've arrived at

julia> promote_type(Irrational{:π}, Bool)
Float64

julia> Base.infer_effects(sin, (Float32,))
(+c,+e,?n,?t,+s)

OTOH (in a fresh session again):

julia> Base.promote_result(Irrational{:π}, Bool, Float64, Irrational{:π})
Float64

julia> promote_type(Irrational{:π}, Bool)
Float64

julia> Base.infer_effects(sin, (Float32,))
(+c,+e,?n,+t,+s)