Error in testset mpfr

[fredrikekre]

From https://build.julialang.org/#/builders/65/builds/674/steps/5/logs/stdio

Error in testset mpfr:
Test Failed at C:\buildbot\worker-tabularasa\tester_win64\build\share\julia\test\mpfr.jl:609
  Expression: log2(x) == log2(42)
   Evaluated: 5.3923174227787607 == 5.39231742277876

[fredrikekre]

Happens on every win64 run.

[fredrikekre]

#41383 (comment)

[aviatesk]

Hmm I'm still not quite sure about the purpose of test (rather expected behaviors of BigFloat in general).

@kshyatt how these precisions are decided and why do we test these equalities under the specific precision ?

julia/test/mpfr.jl

Lines 603 to 616 in d09c8e0

	@testset "log / log2 / log10" begin
	setprecision(53) do
	x = BigFloat(42)
	@test log(x) == log(42)
	@test isinf(log(BigFloat(0)))
	@test_throws DomainError log(BigFloat(-1))
	@test log2(x) == log2(42)
	@test isinf(log2(BigFloat(0)))
	@test_throws DomainError log2(BigFloat(-1))
	@test log10(x) == log10(42)
	@test isinf(log10(BigFloat(0)))
	@test_throws DomainError log10(BigFloat(-1))
	end
	end

[thofma]

Because BigFloat with 53 bits precision should be the same as double:

In particular, with a precision of 53 bits and in any of the four standard rounding modes,
MPFR is able to exactly reproduce all computations with double-precision machine floatingpoint numbers

(See page 2 of https://www.mpfr.org/mpfr-current/mpfr.pdf)

[aviatesk]

Thanks, I see.

With the help from Slack, I confirmed the error is caused by the change of log2(::Float64)'s behavior:

on Windows, before that PR:

julia> versioninfo()
Julia Version 1.8.0-DEV.94
Commit db71a65985 (2021-06-29 21:40 UTC)
Platform Info:
  OS: Windows (x86_64-w64-mingw32)
  CPU: AMD Ryzen 9 3900X 12-Core Processor
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-12.0.0 (ORCJIT, znver2)

julia> log2(42)
  5.392317422778761

julia> setprecision(53) do; log2(BigFloat(42)); end
  5.3923174227787607

log2 internally uses @horner, which is certainly affected by the PR.