J3DShapeDraw and J3DSkinDeform OK (#2908)

configure.py

@@ -1235,7 +1235,7 @@ config.libs = [
             Object(MatchingFor(ALL_GCN), "JSystem/J3DGraphBase/J3DTexture.cpp"),
             Object(MatchingFor(ALL_GCN), "JSystem/J3DGraphBase/J3DPacket.cpp"),
             Object(MatchingFor(ALL_GCN), "JSystem/J3DGraphBase/J3DShapeMtx.cpp"),
-            Object(NonMatching, "JSystem/J3DGraphBase/J3DShapeDraw.cpp"),
+            Object(MatchingFor(ALL_GCN, "ShieldD"), "JSystem/J3DGraphBase/J3DShapeDraw.cpp"),
             Object(MatchingFor(ALL_GCN), "JSystem/J3DGraphBase/J3DShape.cpp"),
             Object(MatchingFor(ALL_GCN), "JSystem/J3DGraphBase/J3DMaterial.cpp"),
             Object(Equivalent, "JSystem/J3DGraphBase/J3DMatBlock.cpp"), # virtual function order
@@ -1254,7 +1254,7 @@ config.libs = [
             Object(MatchingFor(ALL_GCN), "JSystem/J3DGraphAnimator/J3DModel.cpp"),
             Object(MatchingFor(ALL_GCN), "JSystem/J3DGraphAnimator/J3DAnimation.cpp"),
             Object(MatchingFor(ALL_GCN), "JSystem/J3DGraphAnimator/J3DMaterialAnm.cpp"),
-            Object(NonMatching, "JSystem/J3DGraphAnimator/J3DSkinDeform.cpp"),
+            Object(MatchingFor(ALL_GCN), "JSystem/J3DGraphAnimator/J3DSkinDeform.cpp"),
             Object(MatchingFor(ALL_GCN), "JSystem/J3DGraphAnimator/J3DCluster.cpp"),
             Object(MatchingFor(ALL_GCN), "JSystem/J3DGraphAnimator/J3DJoint.cpp"),
             Object(MatchingFor(ALL_GCN), "JSystem/J3DGraphAnimator/J3DMaterialAttach.cpp"),

src/JSystem/J3DGraphAnimator/J3DSkinDeform.cpp

@@ -347,10 +347,6 @@ int J3DSkinDeform::initMtxIndexArray(J3DModelData* pModelData) {
     return kJ3DError_Success;
 }
 
-// NONMATCHING - instruction ordering/optimization issue, matches debug
-// the compiler needs to delay adding +3 to dl until the end of the while loop for the function to match
-// but instead it puts the +3 at the start of the for loop and reworks the other instructions
-// can get a 99.93% match on retail by moving where dl is incremented, but it seems fake as it breaks debug, and introduces an operand swap on src
 void J3DSkinDeform::changeFastSkinDL(J3DModelData* pModelData) {
     J3D_ASSERT_NULLPTR(740, pModelData != NULL);
     for (u16 i = 0; i < pModelData->getShapeNum(); i++) {
@@ -389,7 +385,7 @@ void J3DSkinDeform::changeFastSkinDL(J3DModelData* pModelData) {
                         memcpy(dst, src + 1, (int)(vtxSize - 1)); // The -1 is to remove GX_VA_PNMTXIDX
                         dst += (int)(vtxSize - 1);
                     }
-                    dl += vtxSize * vtxCount;
+                    dl = (u8*)dl + vtxSize * vtxCount;
                 }
 
                 int dlistSize = ((int)dst - (int)displayListStart + 0x1f) & ~0x1f;

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();) {
+    for (u8* dl = dlStart; (dl - dlStart) < getDisplayListSize();) {
-        if (*dl != GX_TRIANGLEFAN && *dl != GX_TRIANGLESTRIP)
+        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 = (u8*)dl + stride * vtxNum;
-        dl += 3;
     }
 
     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;
+        u8 cmd = *(u8*)oldDL;
-        *newDL++ = h;
+        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)
+        int vtxNum = *(u16*)oldDL;
-        s32 vtxNum = *((u16*)(oldDL + 1));
+        oldDL += 2;
-        *((u16*)newDL) = vtxNum;
+        *(u16*)newDL = vtxNum;
         newDL += 2;
 
-        for (s32 i = 0; i < vtxNum; i++) {
+        for (int i = 0; i < vtxNum; i++) {
-            u8* oldDLVtx = &oldDL[stride * i + 3];
+            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 = (u8*)oldDL + stride * vtxNum;
-        oldDL += 3;
     }
 
     u32 realSize = ALIGN_NEXT((uintptr_t)newDL - (uintptr_t)newDLStart, 0x20);

src/JSystem/JMath/JMATrigonometric.cpp

@@ -70,10 +70,10 @@ struct TAsinAcosTable {
 
 namespace JMath {
 
-TSinCosTable<13, f32> sincosTable_;
+TSinCosTable<13, f32> sincosTable_ ATTRIBUTE_ALIGN(32);
 
-TAtanTable atanTable_;
+TAtanTable atanTable_ ATTRIBUTE_ALIGN(32);
 
-TAsinAcosTable asinAcosTable_;
+TAsinAcosTable asinAcosTable_ ATTRIBUTE_ALIGN(32);
 
 }  // namespace JMath