J3DShape / J3DShapeDraw: More OK

2 files changed, 199 insertions, 98 deletions

diff --git a/libs/JSystem/J3DGraphBase/J3DShape.cpp b/libs/JSystem/J3DGraphBase/J3DShape.cpp
index 306362d494..154850d3e3 100644
--- a/libs/JSystem/J3DGraphBase/J3DShape.cpp
+++ b/libs/JSystem/J3DGraphBase/J3DShape.cpp
@@ -58,7 +58,7 @@ extern "C" void _savegpr_29();
 extern "C" void _restgpr_27();
 extern "C" void _restgpr_28();
 extern "C" void _restgpr_29();
-extern "C" extern u8 j3dDefaultViewNo[4 + 4 /* padding */];
+extern "C" extern u32 j3dDefaultViewNo;
 extern "C" u32 sCurrentPipeline__11J3DShapeMtx;
 extern "C" u32 sCurrentScaleFlag__11J3DShapeMtx;
 extern "C" extern u8 struct_804515B0[4];
@@ -72,32 +72,38 @@ enum {
     kVcdVatDLSize = 0xC0,
 };
 
-/* ############################################################################################## */
-/* 804563A0-804563A8 0049A0 0004+04 1/1 0/0 0/0 .sdata2          @687 */
-SECTION_SDATA2 static f32 lit_687[1 + 1 /* padding */] = {
-    0.0f,
-    /* padding */
-    0.0f,
-};
-
-/* 80314B48-80314BB8 30F488 0070+00 0/0 1/1 0/0 .text            initialize__8J3DShapeFv */
-#pragma push
-#pragma optimization_level 0
-#pragma optimizewithasm off
-asm void J3DShape::initialize() {
-    nofralloc
-#include "asm/JSystem/J3DGraphBase/J3DShape/initialize__8J3DShapeFv.s"
+void J3DShape::initialize() { 
+  mMaterial = NULL;
+  mIndex = -1;
+  mMtxGroupNum = 0;
+  mFlags = 0;
+  mRadius = 0.0f;
+  mMin.x = 0.0f;
+  mMin.y = 0.0f;
+  mMin.z = 0.0f;
+  mMax.x = 0.0f;
+  mMax.y = 0.0f;
+  mMax.z = 0.0f;
+  mVtxDesc = NULL;
+  mShapeMtx = NULL;
+  mShapeDraw = NULL;
+  mVertexData = NULL;
+  mDrawMtxData = NULL;
+  mScaleFlagArray = NULL;
+  mDrawMtx = NULL;
+  mNrmMtx = NULL;
+  mCurrentViewNo = &j3dDefaultViewNo;
+  mHasNBT = false;
+  mHasPNMTXIdx = false;
 }
-#pragma pop
 
 /* 80314BB8-80314CBC 30F4F8 0104+00 0/0 1/1 0/0 .text addTexMtxIndexInDL__8J3DShapeF7_GXAttrUl */
-#ifdef NONMATCHING
-void J3DShape::addTexMtxIndexInDL(GXAttr attr, u32 param_1) {
+void J3DShape::addTexMtxIndexInDL(GXAttr attr, u32 valueBase) {
     u32 kSize[] = { 0, 1, 1, 2 };
 
     s32 pnmtxidxOffs = -1;
     s32 attrOffs = -1;
-    u32 stride = 0;
+    s32 stride = 0;
     bool found = false;
 
     for (GXVtxDescList* vtxDesc = mVtxDesc; vtxDesc->attr != GX_VA_NULL; vtxDesc++) {
@@ -115,28 +121,9 @@ void J3DShape::addTexMtxIndexInDL(GXAttr attr, u32 param_1) {
     if (pnmtxidxOffs == -1)
         return;
 
-    for (u16 i = 0; i < mMtxGroupNum; i++)
-        mShapeDraw[i]->addTexMtxIndexInDL(stride, attrOffs, param_1);
-}
-#else
-
-/* ############################################################################################## */
-/* 803A1E98-803A1EA8 02E4F8 0010+00 1/1 0/0 0/0 .rodata          @697 */
-SECTION_RODATA static u8 const lit_697[16] = {
-    0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x02,
-};
-COMPILER_STRIP_GATE(0x803A1E98, &lit_697);
-
-#pragma push
-#pragma optimization_level 0
-#pragma optimizewithasm off
-asm void J3DShape::addTexMtxIndexInDL(_GXAttr param_0, u32 param_1) {
-    nofralloc
-#include "asm/JSystem/J3DGraphBase/J3DShape/addTexMtxIndexInDL__8J3DShapeF7_GXAttrUl.s"
+    for (u16 i = 0; i < getMtxGroupNum(); i++)
+        getShapeDraw(i)->addTexMtxIndexInDL(stride, attrOffs, (s32)valueBase);
 }
-#pragma pop
-
-#endif
 
 /* 80314CBC-80314DA8 30F5FC 00EC+00 0/0 1/1 0/0 .text addTexMtxIndexInVcd__8J3DShapeF7_GXAttr */
 void J3DShape::addTexMtxIndexInVcd(GXAttr attr) {
@@ -177,15 +164,15 @@ void J3DShape::addTexMtxIndexInVcd(GXAttr attr) {
 /* 80314DA8-80314E28 30F6E8 0080+00 0/0 1/1 0/0 .text
  * calcNBTScale__8J3DShapeFRC3VecPA3_A3_fPA3_A3_f               */
 void J3DShape::calcNBTScale(Vec const& param_0, f32 (*param_1)[3][3], f32 (*param_2)[3][3]) {
-    for (u16 i = 0; i < mMtxGroupNum; i++)
-        mShapeMtx[i]->calcNBTScale(param_0, param_1, param_2);
+    for (u16 i = 0; i < getMtxGroupNum(); i++)
+        getShapeMtx(i)->calcNBTScale(param_0, param_1, param_2);
 }
 
 /* 80314E28-80314E98 30F768 0070+00 0/0 1/1 0/0 .text            countBumpMtxNum__8J3DShapeCFv */
 u32 J3DShape::countBumpMtxNum() const {
     u32 num = 0;
-    for (u16 i = 0; i < mMtxGroupNum; i++)
-        num += mShapeMtx[i]->getUseMtxNum();
+    for (u16 i = 0; i < getMtxGroupNum(); i++)
+        num += getShapeMtx(i)->getUseMtxNum();
     return num;
 }
 
@@ -222,14 +209,87 @@ bool J3DShape::isSameVcdVatCmd(J3DShape* other) {
 }
 
 /* 80314F98-80315260 30F8D8 02C8+00 1/1 0/0 0/0 .text            makeVtxArrayCmd__8J3DShapeFv */
-#pragma push
-#pragma optimization_level 0
-#pragma optimizewithasm off
-asm void J3DShape::makeVtxArrayCmd() {
-    nofralloc
-#include "asm/JSystem/J3DGraphBase/J3DShape/makeVtxArrayCmd__8J3DShapeFv.s"
+void J3DShape::makeVtxArrayCmd() {
+    GXVtxAttrFmtList *vtxAttr = mVertexData->getVtxAttrFmtList();
+
+    u8 stride[0x0C];
+    void* array[0x0C];
+    for (u32 i = 0; i < 0x0C; i++) {
+        stride[i] = 0;
+        array[i] = 0;
+    }
+
+    for (; vtxAttr->mAttrib != GX_VA_NULL; vtxAttr++) {
+        switch (vtxAttr->mAttrib) {
+        case GX_VA_POS:
+            {
+                if (vtxAttr->mCompType == GX_F32)
+                    stride[vtxAttr->mAttrib - GX_VA_POS] = 0x0C;
+                else
+                    stride[vtxAttr->mAttrib - GX_VA_POS] = 0x06;
+                array[vtxAttr->mAttrib - GX_VA_POS] = mVertexData->getVtxPosArray();
+                mVertexData->setVtxPosFrac(vtxAttr->mCompShift);
+                mVertexData->setVtxPosType((GXCompType) vtxAttr->mCompType);
+            }
+            break;
+        case GX_VA_NRM:
+            {
+                if (vtxAttr->mCompType == GX_F32)
+                    stride[vtxAttr->mAttrib - GX_VA_POS] = 0x0C;
+                else
+                    stride[vtxAttr->mAttrib - GX_VA_POS] = 0x06;
+                array[vtxAttr->mAttrib - GX_VA_POS] = mVertexData->getVtxNrmArray();
+                mVertexData->setVtxNrmFrac(vtxAttr->mCompShift);
+                mVertexData->setVtxNrmType((GXCompType) vtxAttr->mCompType);
+            }
+            break;
+        case GX_VA_CLR0:
+        case GX_VA_CLR1:
+            {
+                stride[vtxAttr->mAttrib - GX_VA_POS] = 0x04;
+                array[vtxAttr->mAttrib - GX_VA_POS] = mVertexData->getVtxColorArray(vtxAttr->mAttrib - GX_VA_CLR0);
+            }
+            break;
+        case GX_VA_TEX0:
+        case GX_VA_TEX1:
+        case GX_VA_TEX2:
+        case GX_VA_TEX3:
+        case GX_VA_TEX4:
+        case GX_VA_TEX5:
+        case GX_VA_TEX6:
+        case GX_VA_TEX7:
+            {
+                if (vtxAttr->mCompType == GX_F32)
+                    stride[vtxAttr->mAttrib - GX_VA_POS] = 0x08;
+                else
+                    stride[vtxAttr->mAttrib - GX_VA_POS] = 0x04;
+                array[vtxAttr->mAttrib - GX_VA_POS] = mVertexData->getVtxTexCoordArray(vtxAttr->mAttrib - GX_VA_TEX0);
+            }
+            break;
+        default:
+            break;
+        }
+    }
+
+    GXVtxDescList* vtxDesc = mVtxDesc;
+    mHasPNMTXIdx = false;
+    for (; vtxDesc->attr != GX_VA_NULL; vtxDesc++) {
+        if (vtxDesc->attr == GX_VA_NBT && vtxDesc->type != GX_NONE) {
+            mHasNBT = true;
+            stride[GX_VA_NRM - GX_VA_POS] *= 3;
+            array[GX_VA_NRM - GX_VA_POS] = mVertexData->getVtxNBTArray();
+        } else if (vtxDesc->attr == GX_VA_PNMTXIDX && vtxDesc->type != GX_NONE) {
+            mHasPNMTXIdx = true;
+        }
+    }
+
+    for (u32 i = 0; i < 0x0C; i++) {
+        if (array[i] != 0)
+            GDSetArray((GXAttr)(i + GX_VA_POS), array[i], stride[i]);
+        else
+            GDSetArrayRaw((GXAttr)(i + GX_VA_POS), NULL, stride[i]);
+    }
 }
-#pragma pop
 
 /* ############################################################################################## */
 /* 804515C8-804515CC 000AC8 0004+00 1/1 0/0 0/0 .sbss            sInterruptFlag$903 */
@@ -264,14 +324,14 @@ void J3DShape::makeVcdVatCmd() {
 void* J3DShape::sOldVcdVatCmd;
 
 /* 80315300-80315398 30FC40 0098+00 2/2 6/6 3/3 .text            loadPreDrawSetting__8J3DShapeCFv */
-#pragma push
-#pragma optimization_level 0
-#pragma optimizewithasm off
-asm void J3DShape::loadPreDrawSetting() const {
-    nofralloc
-#include "asm/JSystem/J3DGraphBase/J3DShape/loadPreDrawSetting__8J3DShapeCFv.s"
+void J3DShape::loadPreDrawSetting() const {
+    if (sOldVcdVatCmd != mVcdVatCmd) {
+        GXCallDisplayList(mVcdVatCmd, kVcdVatDLSize);
+        sOldVcdVatCmd = mVcdVatCmd;
+    }
+
+    mCurrentMtx.load();
 }
-#pragma pop
 
 /* ############################################################################################## */
 /* 804515D4-804515D8 000AD4 0004+00 3/3 0/0 0/0 .sbss            None */
@@ -308,18 +368,18 @@ void J3DShape::drawFast() const {
         if (J3DShapeMtx::getLODFlag() != 0)
             J3DShapeMtx::resetMtxLoadCache();
 
-        for (u16 n = mMtxGroupNum, i = 0; i < n; i++) {
-            if (mShapeMtx[i] != NULL)
-                mShapeMtx[i]->load();
-            if (mShapeDraw[i] != NULL)
-                mShapeDraw[i]->draw();
+        for (u16 n = getMtxGroupNum(), i = 0; i < n; i++) {
+            if (getShapeMtx(i) != NULL)
+                getShapeMtx(i)->load();
+            if (getShapeDraw(i) != NULL)
+                getShapeDraw(i)->draw();
         }
     } else {
-        J3DFifoLoadPosMtxImm(*j3dSys.mShapePacket->getBaseMtxPtr(), GX_PNMTX0);
-        J3DFifoLoadNrmMtxImm(*j3dSys.mShapePacket->getBaseMtxPtr(), GX_PNMTX0);
-        for (u16 n = mMtxGroupNum, i = 0; i < n; i++)
-            if (mShapeDraw[i] != NULL)
-                mShapeDraw[i]->draw();
+        J3DFifoLoadPosMtxImm(*j3dSys.getShapePacket()->getBaseMtxPtr(), GX_PNMTX0);
+        J3DFifoLoadNrmMtxImm(*j3dSys.getShapePacket()->getBaseMtxPtr(), GX_PNMTX0);
+        for (u16 n = getMtxGroupNum(), i = 0; i < n; i++)
+            if (getShapeDraw(i) != NULL)
+                getShapeDraw(i)->draw();
     }
 }
 #else
@@ -348,9 +408,9 @@ void J3DShape::simpleDraw() const {
     loadPreDrawSetting();
     J3DShapeMtx::setCurrentPipeline(getPipeline());
     loadVtxArray();
-    for (u16 n = mMtxGroupNum, i = 0; i < n; i++) {
-        if (mShapeDraw[i] != NULL) {
-            mShapeDraw[i]->draw();
+    for (u16 n = getMtxGroupNum(), i = 0; i < n; i++) {
+        if (getShapeDraw(i) != NULL) {
+            getShapeDraw(i)->draw();
         }
     }
 }
@@ -366,7 +426,7 @@ void J3DShape::simpleDrawCache() const {
         mCurrentMtx.load();
 
     loadVtxArray();
-    for (u16 n = mMtxGroupNum, i = 0; i < n; i++)
-        if (mShapeDraw[i] != NULL)
-            mShapeDraw[i]->draw();
+    for (u16 n = getMtxGroupNum(), i = 0; i < n; i++)
+        if (getShapeDraw(i) != NULL)
+            getShapeDraw(i)->draw();
 }
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;