J3DShape / J3DShapeDraw: More OK

1 files changed, 64 insertions, 23 deletions

diff --git a/libs/JSystem/J3DGraphBase/J3DShapeDraw.cpp b/libs/JSystem/J3DGraphBase/J3DShapeDraw.cpp
index befab13ce8..6da5f00bcd 100644
--- a/libs/JSystem/J3DGraphBase/J3DShapeDraw.cpp
+++ b/libs/JSystem/J3DGraphBase/J3DShapeDraw.cpp
@@ -4,46 +4,85 @@
 //
 
 #include "JSystem/J3DGraphBase/J3DShapeDraw.h"
+#include "JSystem/JKernel/JKRHeap.h"
 #include "dol2asm.h"
+#include "dolphin/os/OSCache.h"
 #include "dolphin/gx/GX.h"
 #include "dolphin/types.h"
-
-//
-// Forward References:
-//
-
-extern "C" void countVertex__12J3DShapeDrawFUl();
-extern "C" void addTexMtxIndexInDL__12J3DShapeDrawFUlUlUl();
-extern "C" void __ct__12J3DShapeDrawFPCUcUl();
-extern "C" void draw__12J3DShapeDrawCFv();
-extern "C" void __dt__12J3DShapeDrawFv();
+#include "MSL_C/MSL_Common/Src/string.h"
 
 //
 // External References:
 //
 
-SECTION_INIT void memcpy();
+extern "C" void countVertex__12J3DShapeDrawFUl();
 extern "C" void* __nwa__FUli();
 extern "C" void __dl__FPv();
-extern "C" void DCStoreRange();
 extern "C" void _savegpr_18();
 extern "C" void _restgpr_18();
 
-//
-// Declarations:
-//
-
 /* 80314924-80314974 30F264 0050+00 1/1 0/0 0/0 .text            countVertex__12J3DShapeDrawFUl */
-#pragma push
-#pragma optimization_level 0
-#pragma optimizewithasm off
-asm void J3DShapeDraw::countVertex(u32 param_0) {
-    nofralloc
-#include "asm/JSystem/J3DGraphBase/J3DShapeDraw/countVertex__12J3DShapeDrawFUl.s"
+u32 J3DShapeDraw::countVertex(u32 stride) {
+    u32 count = 0;
+    u32 dlStart = (u32)getDisplayList();
+    for (u8* dl = (u8*)dlStart; ((u32)dl - dlStart) < getDisplayListSize();) {
+        if (*dl != 0xA0 && *dl != 0x98)
+            break;
+        u16 vtxNum = *((u16*)(dl + 1));
+        count += vtxNum;
+        dl += stride * vtxNum;
+        dl += 3;
+    }
+    return count;
 }
-#pragma pop
 
+#ifdef NONMATCHING
 /* 80314974-80314ABC 30F2B4 0148+00 0/0 1/1 0/0 .text addTexMtxIndexInDL__12J3DShapeDrawFUlUlUl */
+void J3DShapeDraw::addTexMtxIndexInDL(u32 stride, u32 attrOffs, u32 valueBase) {
+    u32 byteNum = countVertex(stride);
+    u32 newSize = ALIGN_NEXT(mDisplayListSize + byteNum, 0x20);
+    u8* newDLStart = new (0x20) u8[newSize];
+    u8* oldDLStart = getDisplayList();
+    u8* oldDL = oldDLStart;
+    u8* newDL = newDLStart;
+    for (; (oldDL - oldDLStart) < mDisplayListSize;) {
+        // Copy command
+        u8 h = *oldDL;
+        *newDL++ = h;
+
+        if (h != 0xA0 && h != 0x98)
+            break;
+
+        // Copy count
+        // regalloc (I suspect there's a way to shove this in a u16 temp without an mr)
+        u32 vtxNum = *((u16*)(oldDL + 1));
+        *((u16*)newDL) = vtxNum;
+        newDL += 2;
+
+        for (s32 i = 0; i < (u16)vtxNum; i++) {
+            // mul arg swap
+            u8* oldDLVtx = (&oldDL[stride * i + 3]);
+            u8 pnmtxidx = *oldDLVtx;
+            memcpy(newDL, oldDLVtx, attrOffs);
+            u8* newDL1 = &newDL[attrOffs];
+            *newDL1++ = valueBase + pnmtxidx;
+            memcpy(newDL1, oldDLVtx + attrOffs, stride - attrOffs);
+            newDL = newDL1 + (stride - attrOffs);
+        }
+
+        oldDL += stride * vtxNum;
+        oldDL += 3;
+    }
+
+    u32 realSize = ALIGN_NEXT((u32)newDL - (u32)newDLStart, 0x20);
+    for (; (newDL - newDLStart) < newSize; newDL++)
+        *newDL = 0;
+    mDisplayListSize = realSize;
+    mDisplayList = newDLStart;
+    DCStoreRange(newDLStart, mDisplayListSize);
+}
+#else
+
 #pragma push
 #pragma optimization_level 0
 #pragma optimizewithasm off
@@ -53,6 +92,8 @@ asm void J3DShapeDraw::addTexMtxIndexInDL(u32 param_0, u32 param_1, u32 param_2)
 }
 #pragma pop
 
+#endif
+
 /* 80314ABC-80314AD4 30F3FC 0018+00 0/0 1/1 0/0 .text            __ct__12J3DShapeDrawFPCUcUl */
 J3DShapeDraw::J3DShapeDraw(u8 const* displayList, u32 displayListSize) {
     mDisplayList = (void*)displayList;