From a2c55967d4d1102a792df2a6c11c9a9d8d5443b1 Mon Sep 17 00:00:00 2001
From: Hayden Kowalchuk <hayden@hkowsoftware.com>
Date: Wed, 3 Jul 2019 14:30:41 -0400
Subject: [PATCH] fix: possible minor speedups

---
 GL/matrix.c | 8 ++++----
 GL/state.c  | 4 ++--
 Makefile    | 2 +-
 3 files changed, 7 insertions(+), 7 deletions(-)

diff --git a/GL/matrix.c b/GL/matrix.c
index 36f1455..207beca 100644
--- a/GL/matrix.c
+++ b/GL/matrix.c
@@ -50,8 +50,8 @@ void _glInitMatrices() {
     stack_push(&MATRIX_STACKS[1], IDENTITY);
     stack_push(&MATRIX_STACKS[2], IDENTITY);
 
-    memcpy(NORMAL_MATRIX, IDENTITY, sizeof(matrix_t));
-    memcpy(SCREENVIEW_MATRIX, IDENTITY, sizeof(matrix_t));
+    memcpy4(NORMAL_MATRIX, IDENTITY, sizeof(matrix_t));
+    memcpy4(SCREENVIEW_MATRIX, IDENTITY, sizeof(matrix_t));
 
     glDepthRange(0.0f, 1.0f);
     glViewport(0, 0, vid_mode->width, vid_mode->height);
@@ -95,7 +95,7 @@ static void transpose(GLfloat* m) {
 }
 
 static void recalculateNormalMatrix() {
-    memcpy(NORMAL_MATRIX, stack_top(MATRIX_STACKS + (GL_MODELVIEW & 0xF)), sizeof(matrix_t));
+    memcpy4(NORMAL_MATRIX, stack_top(MATRIX_STACKS + (GL_MODELVIEW & 0xF)), sizeof(matrix_t));
     inverse((GLfloat*) NORMAL_MATRIX);
     transpose((GLfloat*) NORMAL_MATRIX);
 }
@@ -225,7 +225,7 @@ void glMultMatrixf(const GLfloat *m) {
         { 0.0f, 0.0f, 0.0f, 1.0f }
     };
 
-    memcpy(TEMP, m, sizeof(matrix_t));
+    memcpy4(TEMP, m, sizeof(matrix_t));
 
     mat_load(stack_top(MATRIX_STACKS + MATRIX_IDX));
     mat_apply(&TEMP);
diff --git a/GL/state.c b/GL/state.c
index 8785f61..0fc5adb 100644
--- a/GL/state.c
+++ b/GL/state.c
@@ -619,10 +619,10 @@ void APIENTRY glGetBooleanv(GLenum pname, GLboolean* params) {
 void APIENTRY glGetFloatv(GLenum pname, GLfloat* params) {
     switch(pname) {
         case GL_PROJECTION_MATRIX:
-            memcpy(params, _glGetProjectionMatrix(), sizeof(float) * 16);
+            memcpy4(params, _glGetProjectionMatrix(), sizeof(float) * 16);
         break;
         case GL_MODELVIEW_MATRIX:
-            memcpy(params, _glGetModelViewMatrix(), sizeof(float) * 16);
+            memcpy4(params, _glGetModelViewMatrix(), sizeof(float) * 16);
         break;
         default:
             _glKosThrowError(GL_INVALID_ENUM, __func__);
diff --git a/Makefile b/Makefile
index fc8b0b0..12ccb21 100644
--- a/Makefile
+++ b/Makefile
@@ -13,7 +13,7 @@ OBJS += containers/stack.o containers/named_array.o containers/aligned_vector.o
 
 SUBDIRS =
 
-KOS_CFLAGS += -ffast-math -Os -Iinclude -funsafe-math-optimizations -fno-expensive-optimizations
+KOS_CFLAGS += -ffast-math -O3 -funroll-loops -Iinclude -funsafe-math-optimizations -fno-expensive-optimizations
 GCC_FLAGS = -mlra
 
 link: