Rebase: Performance regression of scalar randn() between Julia 1.4 an…

…d 1.5 (#39319)

* definition of randn for scalars reverted to 1.4

* Added comments why this change is done

Co-authored-by: Benjamin Lungwitz <52384612+lungben@users.noreply.github.com>

stdlib/Random/src/normal.jl

@@ -35,7 +35,26 @@ julia> randn(rng, ComplexF32, (2, 3))
   0.611224+1.56403im   0.355204-0.365563im  0.0905552+1.31012im
 ```
 """
-@inline randn(rng::AbstractRNG=default_rng()) = _randn(rng, rand(rng, UInt52Raw()))
+@inline function randn(rng::AbstractRNG=default_rng())
+    #=
+    When defining
+    `@inline randn(rng::AbstractRNG=default_rng()) = _randn(rng, rand(rng, UInt52Raw()))`
+    the function call to `_randn` is currently not inlined, resulting in slightly worse
+    performance for scalar random normal numbers than repeating the code of `_randn`
+    inside the following function.
+    =#
+    @inbounds begin
+        r = rand(rng, UInt52Raw())
+
+        # the following code is identical to the one in `_randn(rng::AbstractRNG, r::UInt64)`
+        r &= 0x000fffffffffffff
+        rabs = Int64(r>>1) # One bit for the sign
+        idx = rabs & 0xFF
+        x = ifelse(r % Bool, -rabs, rabs)*wi[idx+1]
+        rabs < ki[idx+1] && return x # 99.3% of the time we return here 1st try
+        return randn_unlikely(rng, idx, rabs, x)
+    end
+end
 
 @inline function _randn(rng::AbstractRNG, r::UInt64)
     @inbounds begin