From 8435ad3f01d6f3d32f95e1fdbe3e83df7ad7595d Mon Sep 17 00:00:00 2001
From: "gregory.hainaut@gmail.com"
 <gregory.hainaut@gmail.com@96395faa-99c1-11dd-bbfe-3dabce05a288>
Date: Sat, 25 Sep 2010 14:01:36 +0000
Subject: [PATCH] GregMiscellaneous: zzogl-pg: * Finish GAS removal. I use SSE2
 instead of mmx because it was easier to write (faster anyway). Code seems
 equivalent between GAS & intrinsics but control flow will need a good testing
 (and some explanation)

Note: I also add a basic C version for reference/debug (not tested yet)


git-svn-id: http://pcsx2.googlecode.com/svn/branches/GregMiscellaneous@3837 96395faa-99c1-11dd-bbfe-3dabce05a288
---
 plugins/zzogl-pg/opengl/zerogs.cpp | 53 ++++++++++++++++++++++++++++++
 1 file changed, 53 insertions(+)

diff --git a/plugins/zzogl-pg/opengl/zerogs.cpp b/plugins/zzogl-pg/opengl/zerogs.cpp
index 88bda55557..081955bb70 100644
--- a/plugins/zzogl-pg/opengl/zerogs.cpp
+++ b/plugins/zzogl-pg/opengl/zerogs.cpp
@@ -32,6 +32,9 @@
 #include "targets.h"
 #include "GLWin.h"
 #include "ZZoglShaders.h"
+#ifdef ZEROGS_SSE2
+#include <emmintrin.h>
+#endif
 
 //----------------------- Defines
 
@@ -851,6 +854,55 @@ bool IsDirty(u32 highdword, u32 psm, int cld, int cbp)
 
 	bool bRet = false;
 
+    // FIXME code generated by intrinsics is the same as the linux asm.
+    // However there is no "cmp %%esi, 0x90" equivalent in the windows asm !!!
+    // So control flow must be check
+#define TEST_THIS
+#ifdef TEST_THIS
+    while(entries != 0) {
+#ifdef ZEROGS_SSE2
+        __m128i result = _mm_cmpeq_epi32(_mm_load_si128((__m128i*)src), _mm_load_si128((__m128i*)dst));
+
+        __m128i result_tmp = _mm_cmpeq_epi32(_mm_load_si128((__m128i*)src+1), _mm_load_si128((__m128i*)dst+1));
+        result = _mm_and_si128(result, result_tmp);
+
+        result_tmp = _mm_cmpeq_epi32(_mm_load_si128((__m128i*)src+2), _mm_load_si128((__m128i*)dst+2));
+        result = _mm_and_si128(result, result_tmp);
+
+        result_tmp = _mm_cmpeq_epi32(_mm_load_si128((__m128i*)src+3), _mm_load_si128((__m128i*)dst+3));
+        result = _mm_and_si128(result, result_tmp);
+
+        u32 result_int = _mm_movemask_epi8(result);
+        if (result_int != 0xFF) {
+            bRet = true;
+            break;
+        }
+#else
+        // I see no point to keep an mmx version. SSE2 versions is probably faster.
+        // Keep a slow portable C version for reference/debug
+        for (int i=0; i < 16 ; i++) {
+            if (*((u32*)src+i) != *((u32*)dst+i)) {
+                bRet = true;
+                break;
+            }
+        }
+#endif
+
+        if (entries & 0x10) {
+            src -= 56; // go back and down one column
+        }
+
+        src += 32; // go to the right block
+
+        if (entries == 0x90) {
+            src += 32; // skip whole block
+        }
+
+        dst += 8;
+        entries -= 16;
+    }
+#else
+
 	// do a fast test with MMX
 #ifdef _MSC_VER
 	int storeebx;
@@ -977,6 +1029,7 @@ Return:
 	".att_syntax\n" : "=m"(bRet) : "c"(dst), "d"(src), "S"(entries) : "eax", "memory");
 
 #endif // _WIN32
+#endif
 	return bRet;
 }