From 9b9c009de9dfd2fb3ad7cb83a9716d9b6ef9911d Mon Sep 17 00:00:00 2001
From: Christoph Oelckers <coelckers@users.noreply.github.com>
Date: Sun, 15 Dec 2019 21:06:41 +0100
Subject: [PATCH] - copybyte is not the same as memcpy.

It should be inlined to use it, though, because homegrown loops for this stuff may have been ok in 1995, but not in 2019.
---
 source/build/include/pragmas.h |  6 ++++++
 source/build/src/engine.cpp    | 22 ++++++++++++++++------
 source/build/src/pragmas.cpp   | 21 +++++++++++++++++++++
 source/rr/src/net.cpp          | 14 +++++++-------
 source/sw/src/sector.cpp       |  2 +-
 5 files changed, 51 insertions(+), 14 deletions(-)

diff --git a/source/build/include/pragmas.h b/source/build/include/pragmas.h
index 527cb03eb..922917289 100644
--- a/source/build/include/pragmas.h
+++ b/source/build/include/pragmas.h
@@ -189,6 +189,9 @@ void qinterpolatedown16short(intptr_t bufptr, int32_t num, int32_t val, int32_t
 #ifndef pragmas_have_clearbuf
 void clearbuf(void *d, int32_t c, int32_t a);
 #endif
+#ifndef pragmas_have_copybuf
+void copybuf(const void *s, void *d, int32_t c);
+#endif
 #ifndef pragmas_have_swaps
 void swapbuf4(void *a, void *b, int32_t c);
 #endif
@@ -196,6 +199,9 @@ void swapbuf4(void *a, void *b, int32_t c);
 #ifndef pragmas_have_clearbufbyte
 void clearbufbyte(void *D, int32_t c, int32_t a);
 #endif
+#ifndef pragmas_have_copybufbyte
+void copybufbyte(const void *S, void *D, int32_t c);
+#endif
 #ifndef pragmas_have_copybufreverse
 void copybufreverse(const void *S, void *D, int32_t c);
 #endif
diff --git a/source/build/src/engine.cpp b/source/build/src/engine.cpp
index 81cdea343..bce1a3be7 100644
--- a/source/build/src/engine.cpp
+++ b/source/build/src/engine.cpp
@@ -4478,7 +4478,7 @@ static void classicDrawBunches(int32_t bunch)
                         smostwall[smostwallcnt] = z;
                         smostwalltype[smostwallcnt] = 1;   //1 for umost
                         smostwallcnt++;
-                        memcpy(&umost[x1],&smost[smostcnt],i*sizeof(smost[0]));
+                        copybufbyte(&umost[x1],&smost[smostcnt],i*sizeof(smost[0]));
                         smostcnt += i;
                     }
                 }
@@ -4564,7 +4564,7 @@ static void classicDrawBunches(int32_t bunch)
                         smostwall[smostwallcnt] = z;
                         smostwalltype[smostwallcnt] = 2;   //2 for dmost
                         smostwallcnt++;
-                        memcpy(&dmost[x1],&smost[smostcnt],i*sizeof(smost[0]));
+                        copybufbyte(&dmost[x1],&smost[smostcnt],i*sizeof(smost[0]));
                         smostcnt += i;
                     }
                 }
@@ -5033,6 +5033,16 @@ static void classicDrawVoxel(int32_t dasprx, int32_t daspry, int32_t dasprz, int
         }
     }
 
+#if 0
+    for (x=0; x<xdimen; x++)
+    {
+        if (daumost[x]>=0 && daumost[x]<ydimen)
+            *(char *)(frameplace + x + bytesperline*daumost[x]) = editorcolors[13];
+        if (dadmost[x]>=0 && dadmost[x]<ydimen)
+            *(char *)(frameplace + x + bytesperline*dadmost[x]) = editorcolors[14];
+    }
+#endif
+
     videoEndDrawing();   //}}}
 }
 
@@ -11638,8 +11648,8 @@ void renderSetTarget(int16_t tilenume, int32_t xsiz, int32_t ysiz)
     rendmode = REND_CLASSIC;
 #endif
 
-    memcpy(&startumost[windowxy1.x],&bakumost[windowxy1.x],(windowxy2.x-windowxy1.x+1)*sizeof(bakumost[0]));
-    memcpy(&startdmost[windowxy1.x],&bakdmost[windowxy1.x],(windowxy2.x-windowxy1.x+1)*sizeof(bakdmost[0]));
+    copybufbyte(&startumost[windowxy1.x],&bakumost[windowxy1.x],(windowxy2.x-windowxy1.x+1)*sizeof(bakumost[0]));
+    copybufbyte(&startdmost[windowxy1.x],&bakdmost[windowxy1.x],(windowxy2.x-windowxy1.x+1)*sizeof(bakdmost[0]));
     setviewcnt++;
 
     offscreenrendering = 1;
@@ -11673,8 +11683,8 @@ void renderRestoreTarget(void)
     ydim = bakysiz[setviewcnt];
     videoSetViewableArea(bakwindowxy1[setviewcnt].x,bakwindowxy1[setviewcnt].y,
             bakwindowxy2[setviewcnt].x,bakwindowxy2[setviewcnt].y);
-    memcpy(&bakumost[windowxy1.x],&startumost[windowxy1.x],(windowxy2.x-windowxy1.x+1)*sizeof(startumost[0]));
-    memcpy(&bakdmost[windowxy1.x],&startdmost[windowxy1.x],(windowxy2.x-windowxy1.x+1)*sizeof(startdmost[0]));
+    copybufbyte(&bakumost[windowxy1.x],&startumost[windowxy1.x],(windowxy2.x-windowxy1.x+1)*sizeof(startumost[0]));
+    copybufbyte(&bakdmost[windowxy1.x],&startdmost[windowxy1.x],(windowxy2.x-windowxy1.x+1)*sizeof(startdmost[0]));
     frameplace = bakframeplace[setviewcnt];
 
     calc_ylookup((setviewcnt == 0) ? bytesperline : bakxsiz[setviewcnt],
diff --git a/source/build/src/pragmas.cpp b/source/build/src/pragmas.cpp
index 1d81f48e6..219849540 100644
--- a/source/build/src/pragmas.cpp
+++ b/source/build/src/pragmas.cpp
@@ -71,6 +71,17 @@ void clearbuf(void *d, int32_t c, int32_t a)
 }
 #endif
 
+#ifndef pragmas_have_copybuf
+void copybuf(const void *s, void *d, int32_t c)
+{
+    auto p = (const int32_t *) s;
+    auto q = (int32_t *) d;
+
+    while (c--)
+        *q++ = *p++;
+}
+#endif
+
 #ifndef pragmas_have_swaps
 void swapbuf4(void *a, void *b, int32_t c)
 {
@@ -102,6 +113,16 @@ void clearbufbyte(void *D, int32_t c, int32_t a)
 }
 #endif
 
+#ifndef pragmas_have_copybufbyte
+void copybufbyte(const void *s, void *d, int32_t c)
+{
+    auto src = (const char *)s;
+    auto dst = (char *)d;
+
+    while (c--)
+        *dst++ = *src++;
+}
+#endif
 
 
 // copybufreverse() is a special case: use the assembly version for GCC on x86
diff --git a/source/rr/src/net.cpp b/source/rr/src/net.cpp
index 9d5838e43..f46f7cd9a 100644
--- a/source/rr/src/net.cpp
+++ b/source/rr/src/net.cpp
@@ -1802,7 +1802,7 @@ void Net_GetInput(void)
     
     if (g_player[myconnectindex].movefifoend&(g_movesPerPacket-1))
     {
-        memcpy(&inputfifo[(g_player[myconnectindex].movefifoend-1)&(MOVEFIFOSIZ-1)][myconnectindex],
+        copybufbyte(&inputfifo[(g_player[myconnectindex].movefifoend-1)&(MOVEFIFOSIZ-1)][myconnectindex],
                     &inputfifo[g_player[myconnectindex].movefifoend&(MOVEFIFOSIZ-1)][myconnectindex],sizeof(input_t));
         g_player[myconnectindex].movefifoend++;
         return;
@@ -2180,7 +2180,7 @@ void Net_ParsePacket(uint8_t *packbuf, int packbufleng)
                 continue;
             }
 
-            memcpy(&osyn[i],&nsyn[i],sizeof(input_t));
+            copybufbyte(&osyn[i],&nsyn[i],sizeof(input_t));
             if (l&1)   nsyn[i].fvel = packbuf[j]+((short)packbuf[j+1]<<8), j += 2;
             if (l&2)   nsyn[i].svel = packbuf[j]+((short)packbuf[j+1]<<8), j += 2;
             if (l&4)
@@ -2219,7 +2219,7 @@ void Net_ParsePacket(uint8_t *packbuf, int packbufleng)
         if (i != myconnectindex)
             for (j=g_movesPerPacket-1;j>=1;j--)
             {
-                memcpy(&nsyn[i],&inputfifo[g_player[i].movefifoend&(MOVEFIFOSIZ-1)][i],sizeof(input_t));
+                copybufbyte(&nsyn[i],&inputfifo[g_player[i].movefifoend&(MOVEFIFOSIZ-1)][i],sizeof(input_t));
                 g_player[i].movefifoend++;
             }
 
@@ -2233,7 +2233,7 @@ void Net_ParsePacket(uint8_t *packbuf, int packbufleng)
         osyn = (input_t *)&inputfifo[(g_player[other].movefifoend-1)&(MOVEFIFOSIZ-1)][0];
         nsyn = (input_t *)&inputfifo[(g_player[other].movefifoend)&(MOVEFIFOSIZ-1)][0];
 
-        memcpy(&osyn[other],&nsyn[other],sizeof(input_t));
+        copybufbyte(&osyn[other],&nsyn[other],sizeof(input_t));
         if (k&1)   nsyn[other].fvel = packbuf[j]+((short)packbuf[j+1]<<8), j += 2;
         if (k&2)   nsyn[other].svel = packbuf[j]+((short)packbuf[j+1]<<8), j += 2;
         if (k&4)
@@ -2272,7 +2272,7 @@ void Net_ParsePacket(uint8_t *packbuf, int packbufleng)
 
         for (i=g_movesPerPacket-1;i>=1;i--)
         {
-            memcpy(&nsyn[other],&inputfifo[g_player[other].movefifoend&(MOVEFIFOSIZ-1)][other],sizeof(input_t));
+            copybufbyte(&nsyn[other],&inputfifo[g_player[other].movefifoend&(MOVEFIFOSIZ-1)][other],sizeof(input_t));
             g_player[other].movefifoend++;
         }
 
@@ -2296,7 +2296,7 @@ void Net_ParsePacket(uint8_t *packbuf, int packbufleng)
         osyn = (input_t *)&inputfifo[(g_player[other].movefifoend-1)&(MOVEFIFOSIZ-1)][0];
         nsyn = (input_t *)&inputfifo[(g_player[other].movefifoend)&(MOVEFIFOSIZ-1)][0];
 
-        memcpy(&osyn[other],&nsyn[other],sizeof(input_t));
+        copybufbyte(&osyn[other],&nsyn[other],sizeof(input_t));
         k = packbuf[j] + (int)(packbuf[j+1]<<8);
         j += 2;
 
@@ -2332,7 +2332,7 @@ void Net_ParsePacket(uint8_t *packbuf, int packbufleng)
 
         for (i=g_movesPerPacket-1;i>=1;i--)
         {
-            memcpy(&nsyn[other],&inputfifo[g_player[other].movefifoend&(MOVEFIFOSIZ-1)][other],sizeof(input_t));
+            copybufbyte(&nsyn[other],&inputfifo[g_player[other].movefifoend&(MOVEFIFOSIZ-1)][other],sizeof(input_t));
             g_player[other].movefifoend++;
         }
 
diff --git a/source/sw/src/sector.cpp b/source/sw/src/sector.cpp
index e62347ed8..d393242e2 100644
--- a/source/sw/src/sector.cpp
+++ b/source/sw/src/sector.cpp
@@ -3178,7 +3178,7 @@ void movelava(char *dapic)
     offs2 = (LAVASIZ + 2) + 1 + ((intptr_t) lavabakpic);
     for (x = 0; x < LAVASIZ; x++)
     {
-        memcpy(offs, offs2, LAVASIZ);
+        copybuf(offs, offs2, LAVASIZ >> 2);
         offs += LAVASIZ;
         offs2 += LAVASIZ + 2;
     }