Add and use SSE-enabled ImRsqrt() in place of 1.0f / ImSqrt(). (#4091)

Squashed 3 commits.

Author: wolfpld

docs/CHANGELOG.txt

@@ -63,6 +63,7 @@ Other Changes:
 - Popups: Added 'OpenPopup(ImGuiID id)' overload to facilitate calling from nested stacks. (#3993, #331) [@zlash]
 - Optimization: Disabling some of MSVC most aggressive Debug runtime checks for some simple/low-level functions
   (e.g. ImVec2, ImVector) leading to a 10-20% increase of performances with MSVC "default" Debug settings.
+- ImDrawList: Add and use SSE-enabled ImRsqrt() in place of 1.0f / ImSqrt(). (#4091) [@wolfpld]
 - ImDrawList: Fixed/improved thickness of thick strokes with sharp angles. (#4053, #3366, #2964, #2868, #2518, #2183)
   Effectively introduced a regression in 1.67 (Jan 2019), and a fix in 1.70 (Apr 2019) but the fix wasn't actually on
   par with original version. Now incorporating the correct revert.

imgui_draw.cpp

@@ -694,7 +694,7 @@ void ImDrawList::PrimQuadUV(const ImVec2& a, const ImVec2& b, const ImVec2& c, c
 
 // On AddPolyline() and AddConvexPolyFilled() we intentionally avoid using ImVec2 and superfluous function calls to optimize debug/non-inlined builds.
 // Those macros expects l-values.
-#define IM_NORMALIZE2F_OVER_ZERO(VX,VY)     do { float d2 = VX*VX + VY*VY; if (d2 > 0.0f) { float inv_len = 1.0f / ImSqrt(d2); VX *= inv_len; VY *= inv_len; } } while (0)
+#define IM_NORMALIZE2F_OVER_ZERO(VX,VY)     do { float d2 = VX*VX + VY*VY; if (d2 > 0.0f) { float inv_len = ImRsqrt(d2); VX *= inv_len; VY *= inv_len; } } while (0)
 #define IM_FIXNORMAL2F_MAX_INVLEN2          100.0f // 500.0f (see #4053, #3366)
 #define IM_FIXNORMAL2F(VX,VY)               do { float d2 = VX*VX + VY*VY; if (d2 > 0.000001f) { float inv_len2 = 1.0f / d2; if (inv_len2 > IM_FIXNORMAL2F_MAX_INVLEN2) inv_len2 = IM_FIXNORMAL2F_MAX_INVLEN2; VX *= inv_len2; VY *= inv_len2; } } while (0)
 

imgui_internal.h

@@ -51,6 +51,12 @@ Index of this file:
 #include <math.h>       // sqrtf, fabsf, fmodf, powf, floorf, ceilf, cosf, sinf
 #include <limits.h>     // INT_MIN, INT_MAX
 
+// Enable SSE intrinsics if available
+#if defined __SSE__ || defined __x86_64__ || defined _M_X64
+#define IMGUI_ENABLE_SSE
+#include <immintrin.h>
+#endif
+
 // Visual Studio warnings
 #ifdef _MSC_VER
 #pragma warning (push)
@@ -390,6 +396,12 @@ static inline float  ImAbs(float x)             { return fabsf(x); }
 static inline double ImAbs(double x)            { return fabs(x); }
 static inline float  ImSign(float x)            { return (x < 0.0f) ? -1.0f : ((x > 0.0f) ? 1.0f : 0.0f); } // Sign operator - returns -1, 0 or 1 based on sign of argument
 static inline double ImSign(double x)           { return (x < 0.0) ? -1.0 : ((x > 0.0) ? 1.0 : 0.0); }
+#ifdef IMGUI_ENABLE_SSE
+static inline float  ImRsqrt(float x)           { return _mm_cvtss_f32(_mm_rsqrt_ss(_mm_set_ss(x))); }
+#else
+static inline float  ImRsqrt(float x)           { return 1.0f / sqrtf(x); }
+#endif
+static inline double ImRsqrt(double x)          { return 1.0 / sqrt(x); }
 #endif
 // - ImMin/ImMax/ImClamp/ImLerp/ImSwap are used by widgets which support variety of types: signed/unsigned int/long long float/double
 // (Exceptionally using templates here but we could also redefine them for those types)
@@ -410,7 +422,7 @@ static inline ImVec4 ImLerp(const ImVec4& a, const ImVec4& b, float t)
 static inline float  ImSaturate(float f)                                        { return (f < 0.0f) ? 0.0f : (f > 1.0f) ? 1.0f : f; }
 static inline float  ImLengthSqr(const ImVec2& lhs)                             { return (lhs.x * lhs.x) + (lhs.y * lhs.y); }
 static inline float  ImLengthSqr(const ImVec4& lhs)                             { return (lhs.x * lhs.x) + (lhs.y * lhs.y) + (lhs.z * lhs.z) + (lhs.w * lhs.w); }
-static inline float  ImInvLength(const ImVec2& lhs, float fail_value)           { float d = (lhs.x * lhs.x) + (lhs.y * lhs.y); if (d > 0.0f) return 1.0f / ImSqrt(d); return fail_value; }
+static inline float  ImInvLength(const ImVec2& lhs, float fail_value)           { float d = (lhs.x * lhs.x) + (lhs.y * lhs.y); if (d > 0.0f) return ImRsqrt(d); return fail_value; }
 static inline float  ImFloor(float f)                                           { return (float)(int)(f); }
 static inline float  ImFloorSigned(float f)                                     { return (float)((f >= 0 || (int)f == f) ? (int)f : (int)f - 1); } // Decent replacement for floorf()
 static inline ImVec2 ImFloor(const ImVec2& v)                                   { return ImVec2((float)(int)(v.x), (float)(int)(v.y)); }