fix rounding mode in construction of BigFloat from pi (#55911)

The default argument of the method was outdated, reading the global
default rounding directly, bypassing the `ScopedValue` stuff.

base/irrationals.jl

@@ -216,7 +216,7 @@ function irrational(sym, val, def)
     esym = esc(sym)
     qsym = esc(Expr(:quote, sym))
     bigconvert = isa(def,Symbol) ? quote
-        function Base.BigFloat(::Irrational{$qsym}, r::MPFR.MPFRRoundingMode=MPFR.ROUNDING_MODE[]; precision=precision(BigFloat))
+        function Base.BigFloat(::Irrational{$qsym}, r::MPFR.MPFRRoundingMode=Rounding.rounding_raw(BigFloat); precision=precision(BigFloat))
             c = BigFloat(;precision=precision)
             ccall(($(string("mpfr_const_", def)), :libmpfr),
                   Cint, (Ref{BigFloat}, MPFR.MPFRRoundingMode), c, r)

test/rounding.jl

@@ -470,3 +470,28 @@ end
         @test prevfloat(f) < i
     end
 end
+
+@testset "π to `BigFloat` with `setrounding`" begin
+    function irrational_to_big_float(c::AbstractIrrational)
+        BigFloat(c)
+    end
+
+    function irrational_to_big_float_with_rounding_mode(c::AbstractIrrational, rm::RoundingMode)
+        f = () -> irrational_to_big_float(c)
+        setrounding(f, BigFloat, rm)
+    end
+
+    function irrational_to_big_float_with_rounding_mode_and_precision(c::AbstractIrrational, rm::RoundingMode, prec::Int)
+        f = () -> irrational_to_big_float_with_rounding_mode(c, rm)
+        setprecision(f, BigFloat, prec)
+    end
+
+    for c ∈ (π, MathConstants.γ, MathConstants.catalan)
+        for p ∈ 1:40
+            @test (
+                irrational_to_big_float_with_rounding_mode_and_precision(c, RoundDown, p) < c <
+                irrational_to_big_float_with_rounding_mode_and_precision(c, RoundUp, p)
+            )
+        end
+    end
+end