From 585a0b464cf8ce81b2c779c7223a7f8163726169 Mon Sep 17 00:00:00 2001
From: Danny Robson <danny@nerdcruft.net>
Date: Fri, 20 Mar 2015 01:32:56 +1100
Subject: [PATCH] maths: make sign query inline

code which needs this function tends to require a fairly simple
implementation inline for the optimiser to reach more successfully. we
tended to generate function calls to this which slowed this inner loops.
---
 maths.cpp | 25 -------------------------
 maths.hpp |  4 +++-
 maths.ipp | 12 ++++++++++++
 3 files changed, 15 insertions(+), 26 deletions(-)

diff --git a/maths.cpp b/maths.cpp
index 9e89ff88..2afc319d 100644
--- a/maths.cpp
+++ b/maths.cpp
@@ -152,31 +152,6 @@ template uint32_t round_pow2 (uint32_t);
 template uint64_t round_pow2 (uint64_t);
 
 
-//-----------------------------------------------------------------------------
-template <typename T>
-int
-sign (T val) {
-    return val >= 0 ? 1 : -1;
-}
-
-
-template <>
-int
-sign (float val) {
-    return static_cast<int> (copysign (1.0f, val));
-}
-
-
-template <>
-int
-sign (double val) {
-    return static_cast<int> (copysign (1.0, val));
-}
-
-
-template int sign (int);
-
-
 //-----------------------------------------------------------------------------
 // Simple instantiations. Some functions aren't used internally to the library
 // so it's easier to instantiate early and check for broken code at library
diff --git a/maths.hpp b/maths.hpp
index b5bbf50e..5f0e0adf 100644
--- a/maths.hpp
+++ b/maths.hpp
@@ -21,6 +21,7 @@
 #define __MATHS_HPP
 
 #include "debug.hpp"
+#include "types/traits.hpp"
 
 #include <cstdint>
 #include <type_traits>
@@ -104,7 +105,8 @@ digits [[gnu::pure]] (const T& value);
 
 //-----------------------------------------------------------------------------
 template <typename T>
-int sign [[gnu::pure]] (T val);
+typename try_signed<T>::type
+sign [[gnu::pure]] (T val);
 
 
 //-----------------------------------------------------------------------------
diff --git a/maths.ipp b/maths.ipp
index 482aaacc..207113ff 100644
--- a/maths.ipp
+++ b/maths.ipp
@@ -45,6 +45,18 @@ pow (T x, unsigned y)
 }
 
 
+///----------------------------------------------------------------------------
+/// Return a unit type with a sign that matches the provided value
+///
+/// This really needs to be inline for performance as it's used in a few inner
+/// loops where the possibility of a function call is a deal breaker.
+template <typename T>
+typename try_signed<T>::type
+sign (T val) {
+    return val >= 0 ? T{1} : T{-1};
+}
+
+
 //-----------------------------------------------------------------------------
 template <>
 struct constants<float>