restore fallback 3-arg setprecision method (#58586)

fixes #55899

---------

Co-authored-by: Sukera <11753998+Seelengrab@users.noreply.github.com>

Author: JeffBezanson

HISTORY.md

@@ -134,6 +134,8 @@ New library features
 * `Timer` now has readable `timeout` and `interval` properties, and a more descriptive `show` method ([#57081]).
 * `sort` now supports `NTuple`s ([#54494]).
 * `map!(f, A)` now stores the results in `A`, like `map!(f, A, A)` or `A .= f.(A)` ([#40632]).
+* `setprecision` with a function argument (typically a `do` block) is now thread safe. Other forms
+  should be avoided, and types should switch to an implementation using `ScopedValue` ([#51362]).
 
 Standard library changes
 ------------------------

base/mpfr.jl

@@ -1183,9 +1183,29 @@ Often used as `setprecision(T, precision) do ... end`
 Note: `nextfloat()`, `prevfloat()` do not use the precision mentioned by
 `setprecision`.
 
+!!! warning
+    There is a fallback implementation of this method that calls `precision`
+    and `setprecision`, but it should no longer be relied on. Instead, you
+    should define the 3-argument form directly in a way that uses `ScopedValue`,
+    or recommend that callers use `ScopedValue` and `@with` themselves.
+
 !!! compat "Julia 1.8"
     The `base` keyword requires at least Julia 1.8.
 """
+function setprecision(f::Function, ::Type{T}, prec::Integer; kws...) where T
+    depwarn("""
+            The fallback `setprecision(::Function, ...)` method is deprecated. Packages overloading this method should
+            implement their own specialization using `ScopedValue` instead.
+            """, :setprecision)
+    old_prec = precision(T)
+    setprecision(T, prec; kws...)
+    try
+        return f()
+    finally
+        setprecision(T, old_prec)
+    end
+end
+
 function setprecision(f::Function, ::Type{BigFloat}, prec::Integer; base::Integer=2)
     Base.ScopedValues.@with(CURRENT_PRECISION => _convert_precision_from_base(prec, base), f())
 end