From 5ea4313d598fd423fa18ad59e0244b2e4fe9f5d8 Mon Sep 17 00:00:00 2001
From: Luke Benstead <kazade@gmail.com>
Date: Fri, 14 May 2021 21:53:27 +0100
Subject: [PATCH] Fix software renderer

---
 GL/draw.c               | 10 +++++-----
 GL/flush.c              |  4 ++--
 GL/platforms/sh4.h      |  2 ++
 GL/platforms/software.h |  2 ++
 4 files changed, 11 insertions(+), 7 deletions(-)

diff --git a/GL/draw.c b/GL/draw.c
index 2054932..99a668d 100644
--- a/GL/draw.c
+++ b/GL/draw.c
@@ -471,7 +471,7 @@ GL_FORCE_INLINE void genQuads(Vertex* output, GLuint count) {
     Vertex* final = output + 3;
     GLuint i = count >> 2;
     while(i--) {
-        __asm__("pref @%0" : : "r"(pen + 4));
+        PREFETCH(pen + 4);
 
         swapVertex(pen, final);
         final->flags = GPU_CMD_VERTEX_EOL;
@@ -639,7 +639,7 @@ static void _readPositionData(ReadDiffuseFunc func, const GLuint first, const GL
     uint32_t* flags;
 
     ITERATE(count) {
-        __asm__("pref @%0" : : "r"(vptr + vstride));
+        PREFETCH(vptr + vstride);
 
         func(vptr, out);
         vptr += vstride;
@@ -660,7 +660,7 @@ static void _readUVData(ReadUVFunc func, const GLuint first, const GLuint count,
     GLubyte* out = (GLubyte*) output[0].uv;
 
     ITERATE(count) {
-        __asm__("pref @%0" : : "r"(uvptr + uvstride));
+        PREFETCH(uvptr + uvstride);
 
         func(uvptr, out);
         uvptr += uvstride;
@@ -675,7 +675,7 @@ static void _readSTData(ReadUVFunc func, const GLuint first, const GLuint count,
     GLubyte* out = (GLubyte*) extra[0].st;
 
     ITERATE(count) {
-        __asm__("pref @%0" : : "r"(stptr + ststride));
+        PREFETCH(stptr + ststride);
 
         func(stptr, out);
         stptr += ststride;
@@ -725,7 +725,7 @@ static void _readDiffuseData(ReadDiffuseFunc func, const GLuint first, const GLu
     GLubyte* out = (GLubyte*) output[0].bgra;
 
     ITERATE(count) {
-        __asm__("pref @%0" : : "r"(cptr + cstride));
+        PREFETCH(cptr + cstride);
 
         func(cptr, out);
         cptr += cstride;
diff --git a/GL/flush.c b/GL/flush.c
index c52b1be..b29aac9 100644
--- a/GL/flush.c
+++ b/GL/flush.c
@@ -101,7 +101,7 @@ GL_FORCE_INLINE void glPerspectiveDivideStandard(void* src, uint32_t n) {
     const float h = GetVideoMode()->height;
 
     while(n--) {
-        __asm__("pref @%0" : : "r"(vertex + 1));
+        PREFETCH(vertex + 1);
 
         if(likely(glIsVertex(vertex->flags))) {
             const float f = MATH_Fast_Invert(vertex->w);
@@ -135,7 +135,7 @@ GL_FORCE_INLINE void glPerspectiveDivideFastMode(void* src, uint32_t n) {
     const float h = GetVideoMode()->height;
 
     while(n--) {
-        __asm__("pref @%0" : : "r"(vertex + 1));
+        PREFETCH(vertex + 1);
 
         if(likely(glIsVertex(vertex->flags))) {
             const float f = MATH_Fast_Invert(vertex->w);
diff --git a/GL/platforms/sh4.h b/GL/platforms/sh4.h
index 0c312be..f2a3295 100644
--- a/GL/platforms/sh4.h
+++ b/GL/platforms/sh4.h
@@ -17,6 +17,8 @@
 #endif
 
 
+#define PREFETCH(addr) __asm__("pref @%0" : : "r"((addr)))
+
 /* We use sq_cpy if the src and size is properly aligned. We control that the
  * destination is properly aligned so we assert that. */
 #define FASTCPY(dst, src, bytes) \
diff --git a/GL/platforms/software.h b/GL/platforms/software.h
index e3a3a03..bbb4f35 100644
--- a/GL/platforms/software.h
+++ b/GL/platforms/software.h
@@ -5,6 +5,8 @@
 
 #include "../types.h"
 
+#define PREFETCH(addr) do {} while(0)
+
 #define MATH_Fast_Divide(n, d) (n / d)
 #define MATH_fmac(a, b, c) (a * b + c)
 #define MATH_Fast_Sqrt(x) sqrtf((x))