J3DShapeDraw and J3DSkinDeform OK (#2908)

1 files changed, 22 insertions, 20 deletions

diff --git a/src/JSystem/J3DGraphBase/J3DShapeDraw.cpp b/src/JSystem/J3DGraphBase/J3DShapeDraw.cpp
index fcfec6e2a4..dc00a46f9a 100644
--- a/src/JSystem/J3DGraphBase/J3DShapeDraw.cpp
+++ b/src/JSystem/J3DGraphBase/J3DShapeDraw.cpp
@@ -8,55 +8,57 @@
 
 u32 J3DShapeDraw::countVertex(u32 stride) {
     u32 count = 0;
-    uintptr_t dlStart = (uintptr_t)getDisplayList();
+    u8* dlStart = (u8*)getDisplayList();
 
-    for (u8* dl = (u8*)dlStart; ((uintptr_t)dl - dlStart) < getDisplayListSize();) {
-        if (*dl != GX_TRIANGLEFAN && *dl != GX_TRIANGLESTRIP)
+    for (u8* dl = dlStart; (dl - dlStart) < getDisplayListSize();) {
+        u8 cmd = *(u8*)dl;
+        dl++;
+        if (cmd != GX_TRIANGLEFAN && cmd != GX_TRIANGLESTRIP)
             break;
-        u16 vtxNum = *((u16*)(dl + 1));
+        int vtxNum = *((u16*)(dl));
+        dl += 2;
         count += vtxNum;
-        dl += stride * vtxNum;
-        dl += 3;
+        dl = (u8*)dl + stride * vtxNum;
     }
 
     return count;
 }
 
-// NONMATCHING regalloc
 void J3DShapeDraw::addTexMtxIndexInDL(u32 stride, u32 attrOffs, u32 valueBase) {
     u32 byteNum = countVertex(stride);
-    u32 newSize = ALIGN_NEXT(mDisplayListSize + byteNum, 0x20);
+    u32 oldSize = mDisplayListSize;
+    u32 newSize = ALIGN_NEXT(oldSize + byteNum, 0x20);
     u8* newDLStart = new (0x20) u8[newSize];
-    u8* oldDLStart = getDisplayList();
+    u8* oldDLStart = (u8*)mDisplayList;
     u8* oldDL = oldDLStart;
     u8* newDL = newDLStart;
 
     for (; (oldDL - oldDLStart) < mDisplayListSize;) {
         // Copy command
-        u8 h = *oldDL;
-        *newDL++ = h;
+        u8 cmd = *(u8*)oldDL;
+        oldDL++;
+        *newDL++ = cmd;
 
-        if (h != GX_TRIANGLEFAN && h != GX_TRIANGLESTRIP)
+        if (cmd != GX_TRIANGLEFAN && cmd != GX_TRIANGLESTRIP)
             break;
 
         // Copy count
-        // regalloc (I suspect there's a way to shove this in a u16 temp without an mr)
-        s32 vtxNum = *((u16*)(oldDL + 1));
-        *((u16*)newDL) = vtxNum;
+        int vtxNum = *(u16*)oldDL;
+        oldDL += 2;
+        *(u16*)newDL = vtxNum;
         newDL += 2;
 
-        for (s32 i = 0; i < vtxNum; i++) {
-            u8* oldDLVtx = &oldDL[stride * i + 3];
+        for (int i = 0; i < vtxNum; i++) {
+            u8* oldDLVtx = &oldDL[stride * i];
             u8 pnmtxidx = *oldDLVtx;
-            memcpy(newDL, oldDLVtx, attrOffs);
+            memcpy(newDL, oldDLVtx, (int)attrOffs);
             newDL += attrOffs;
             *newDL++ = valueBase + pnmtxidx;
             memcpy(newDL, oldDLVtx + attrOffs, stride - attrOffs);
             newDL += (stride - attrOffs);
         }
 
-        oldDL += stride * vtxNum;
-        oldDL += 3;
+        oldDL = (u8*)oldDL + stride * vtxNum;
     }
 
     u32 realSize = ALIGN_NEXT((uintptr_t)newDL - (uintptr_t)newDLStart, 0x20);