Use libquadmath sincos rather than sin and cos (#119)

On macOS AArch64 with libquadmath, this provides a performance
improvement: median 477 ns on this branch vs. 880 ns on master. On
FreeBSD AArch64 with libm, the performance on this branch is identical
to master, about 610 ns for both. (Note that the Julia version used on
macOS and FreeBSD here differed: 1.12.7 on macOS and 1.14.0-DEV.3218 on
FreeBSD.)

Author: ararslan

src/Quadmath.jl

@@ -367,7 +367,12 @@ if VERSION < v"1.14.0-DEV.2070"
     end
 end
 
-sincos(x::Float128) = (sin(x), cos(x))
+function sincos(x::Float128)
+    sin = Ref{Cfloat128}()
+    cos = Ref{Cfloat128}()
+    @quad_ccall(@autoql(sincos)(x::Cfloat128, sin::Ref{Cfloat128}, cos::Ref{Cfloat128})::Cvoid)
+    return (Float128(sin[]), Float128(cos[]))
+end
 
 ## misc
 

test/runtests.jl

@@ -143,6 +143,9 @@ end
         if !Sys.iswindows()
             @test fma(x,x,Float128(-1.0)) ≈ Float128(1)
         end
+        @test sin(x) ≈ sin(xd)
+        @test cos(x) ≈ cos(xd)
+        @test all(sincos(x) .≈ (sin(x), cos(x)))
     end
     @testset "complex" begin
         x = sqrt(ComplexF128(1.0 + 1.0im))