X-Git-Url: https://notaz.gp2x.de/cgi-bin/gitweb.cgi?a=blobdiff_plain;f=plugins%2Fdfxvideo%2Fgpulib_if.c;h=86cfd26844718128b347b39d3af16475d01d6767;hb=4ad17db3c1b6bfa042832d613369eda9c7ffff4f;hp=ff0c96c7092a02e98e7e8218565ebcbc80e8b8db;hpb=857275a93c8e222bc800999f08d4b42e327f4c1b;p=pcsx_rearmed.git

diff --git a/plugins/dfxvideo/gpulib_if.c b/plugins/dfxvideo/gpulib_if.c
index ff0c96c7..86cfd268 100644
--- a/plugins/dfxvideo/gpulib_if.c
+++ b/plugins/dfxvideo/gpulib_if.c
@@ -16,6 +16,7 @@
 #include <stdlib.h>
 #include <string.h>
 #include "../gpulib/gpu.h"
+#include "../../include/arm_features.h"
 
 #define u32 uint32_t
 
@@ -37,10 +38,10 @@
 
 // byteswappings
 
-#define SWAP16(x) ({ uint16_t y=(x); (((y)>>8 & 0xff) | ((y)<<8 & 0xff00)); })
-#define SWAP32(x) ({ uint32_t y=(x); (((y)>>24 & 0xfful) | ((y)>>8 & 0xff00ul) | ((y)<<8 & 0xff0000ul) | ((y)<<24 & 0xff000000ul)); })
+#define SWAP16(x) __builtin_bswap16(x)
+#define SWAP32(x) __builtin_bswap32(x)
 
-#ifdef __BIG_ENDIAN__
+#if __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__
 
 // big endian config
 #define HOST2LE32(x) SWAP32(x)
@@ -68,16 +69,15 @@
 
 #endif
 
-#define GETLEs16(X) ((int16_t)GETLE16((uint16_t *)X))
-#define GETLEs32(X) ((int16_t)GETLE32((uint16_t *)X))
+#define GETLEs16(X) ((int16_t)GETLE16((uint16_t *)(X)))
 
-#define GETLE16(X) LE2HOST16(*(uint16_t *)X)
-#define GETLE32_(X) LE2HOST32(*(uint32_t *)X)
-#define GETLE16D(X) ({uint32_t val = GETLE32(X); (val<<16 | val >> 16);})
-#define PUTLE16(X, Y) do{*((uint16_t *)X)=HOST2LE16((uint16_t)Y);}while(0)
-#define PUTLE32_(X, Y) do{*((uint32_t *)X)=HOST2LE16((uint32_t)Y);}while(0)
-#ifdef __arm__
-#define GETLE32(X) (*(uint16_t *)(X)|(((uint16_t *)(X))[1]<<16))
+#define GETLE16(X) LE2HOST16(*(uint16_t *)(X))
+#define GETLE32_(X) LE2HOST32(*(uint32_t *)(X))
+#define PUTLE16(X, Y) do{*((uint16_t *)(X))=HOST2LE16((uint16_t)(Y));}while(0)
+#define PUTLE32_(X, Y) do{*((uint32_t *)(X))=HOST2LE32((uint32_t)(Y));}while(0)
+#if defined(__arm__) && !defined(HAVE_ARMV6)
+// for (very) old ARMs with no unaligned loads?
+#define GETLE32(X) (*(uint16_t *)(X)|((uint32_t)((uint16_t *)(X))[1]<<16))
 #define PUTLE32(X, Y) do{uint16_t *p_=(uint16_t *)(X);uint32_t y_=Y;p_[0]=y_;p_[1]=y_>>16;}while(0)
 #else
 #define GETLE32 GETLE32_
@@ -219,18 +219,12 @@ extern int32_t           drawH;
 #define KEY_BADTEXTURES   128
 #define KEY_CHECKTHISOUT  256
 
-#if !defined(__BIG_ENDIAN__) || defined(__x86_64__) || defined(__i386__)
-#ifndef __LITTLE_ENDIAN__
-#define __LITTLE_ENDIAN__
-#endif
-#endif
-
-#ifdef __LITTLE_ENDIAN__
+#if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
 #define RED(x) (x & 0xff)
 #define BLUE(x) ((x>>16) & 0xff)
 #define GREEN(x) ((x>>8) & 0xff)
 #define COLOR(x) (x & 0xffffff)
-#elif defined __BIG_ENDIAN__
+#else
 #define RED(x) ((x>>24) & 0xff)
 #define BLUE(x) ((x>>8) & 0xff)
 #define GREEN(x) ((x>>16) & 0xff)
@@ -239,12 +233,8 @@ extern int32_t           drawH;
 
 PSXDisplay_t      PSXDisplay;
 unsigned char  *psxVub;
-signed   char  *psxVsb;
 unsigned short *psxVuw;
 unsigned short *psxVuw_eom;
-signed   short *psxVsw;
-uint32_t *psxVul;
-int32_t  *psxVsl;
 
 long              lGPUstatusRet;
 uint32_t          lGPUInfoVals[16];
@@ -272,13 +262,7 @@ long           lLowerpart;
 static void set_vram(void *vram)
 {
  psxVub=vram;
-
- psxVsb=(signed char *)psxVub;                         // different ways of accessing PSX VRAM
- psxVsw=(signed short *)psxVub;
- psxVsl=(int32_t *)psxVub;
  psxVuw=(unsigned short *)psxVub;
- psxVul=(uint32_t *)psxVub;
-
  psxVuw_eom=psxVuw+1024*512;                           // pre-calc of end of vram
 }
 
@@ -321,7 +305,7 @@ int do_cmd_list(unsigned int *list, int list_len, int *last_cmd)
 
   for (; list < list_end; list += 1 + len)
   {
-    cmd = *list >> 24;
+    cmd = GETLE32(list) >> 24;
     len = cmd_lengths[cmd];
     if (list + 1 + len > list_end) {
       cmd = -1;
@@ -332,7 +316,7 @@ int do_cmd_list(unsigned int *list, int list_len, int *last_cmd)
     if (cmd == 0xa0 || cmd == 0xc0)
       break; // image i/o, forward to upper layer
     else if ((cmd & 0xf8) == 0xe0)
-      gpu.ex_regs[cmd & 7] = list[0];
+      gpu.ex_regs[cmd & 7] = GETLE32(list);
 #endif
 
     primTableJ[cmd]((void *)list);
@@ -351,7 +335,7 @@ int do_cmd_list(unsigned int *list, int list_len, int *last_cmd)
             goto breakloop;
           }
 
-          if((*list_position & 0xf000f000) == 0x50005000)
+          if((*list_position & HOST2LE32(0xf000f000)) == HOST2LE32(0x50005000))
             break;
 
           list_position++;
@@ -374,7 +358,7 @@ int do_cmd_list(unsigned int *list, int list_len, int *last_cmd)
             goto breakloop;
           }
 
-          if((*list_position & 0xf000f000) == 0x50005000)
+          if((*list_position & HOST2LE32(0xf000f000)) == HOST2LE32(0x50005000))
             break;
 
           list_position += 2;
@@ -389,8 +373,8 @@ int do_cmd_list(unsigned int *list, int list_len, int *last_cmd)
       case 0xA0:          //  sys -> vid
       {
         short *slist = (void *)list;
-        u32 load_width = slist[4];
-        u32 load_height = slist[5];
+        u32 load_width = LE2HOST32(slist[4]);
+        u32 load_height = LE2HOST32(slist[5]);
         u32 load_size = load_width * load_height;
 
         len += load_size / 2;
@@ -408,8 +392,16 @@ breakloop:
   return list - list_start;
 }
 
-void renderer_sync_ecmds(uint32_t *ecmds)
+void renderer_sync_ecmds(uint32_t *ecmds_)
 {
+#if __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__
+  // the funcs below expect LE
+  uint32_t i, ecmds[8];
+  for (i = 1; i <= 6; i++)
+    ecmds[i] = HTOLE32(ecmds_[i]);
+#else
+  uint32_t *ecmds = ecmds_;
+#endif
   cmdTexturePage((unsigned char *)&ecmds[1]);
   cmdTextureWindow((unsigned char *)&ecmds[2]);
   cmdDrawAreaStart((unsigned char *)&ecmds[3]);