Copy J3DMatblock inlines from TP, equivalent with 2 remaining regalloc

configure.py

@@ -1004,7 +1004,7 @@ config.libs = [
             Object(Matching,    "JSystem/J3DGraphBase/J3DShapeMtx.cpp", extra_cflags=['-pragma "nosyminline on"']), # TODO: nosyminline may be a fakematch. jsystem pch?
             Object(Matching,    "JSystem/J3DGraphBase/J3DShape.cpp"),
             Object(Matching,    "JSystem/J3DGraphBase/J3DMaterial.cpp"),
-            Object(NonMatching, "JSystem/J3DGraphBase/J3DMatBlock.cpp"),
+            Object(Equivalent,  "JSystem/J3DGraphBase/J3DMatBlock.cpp"), # regalloc
             Object(Matching,    "JSystem/J3DGraphBase/J3DTevs.cpp"),
             Object(Matching,    "JSystem/J3DGraphBase/J3DDrawBuffer.cpp"),
         ],

include/JSystem/J3DGraphBase/J3DMatBlock.h

@@ -854,8 +854,8 @@ struct J3DIndTexMtx : public J3DIndTexMtxInfo {
     explicit J3DIndTexMtx(const J3DIndTexMtxInfo& info) { *(J3DIndTexMtxInfo*)this = info; }
     ~J3DIndTexMtx() {}
 
-    void load(u32 id) {
-        J3DGDSetIndTexMtx(GXIndTexMtxID(id + GX_ITM_0), getOffsetMtx(), getScaleExp());
+    void load(u32 id) const {
+        J3DGDSetIndTexMtx(GXIndTexMtxID(id + GX_ITM_0), (Mtx3P)mOffsetMtx, mScaleExp);
     }
 
     Mtx3P getOffsetMtx() {
@@ -960,17 +960,11 @@ inline u16 calcColorChanID(u16 enable, u8 matSrc, u8 lightMask, u8 diffuseFn, u8
     return reg;
 }
 
-inline u32 setChanCtrlMacro(u8 enable, GXColorSrc ambSrc, GXColorSrc matSrc, u32 lightMask, GXDiffuseFn diffuseFn, GXAttnFn attnFn) {
-    u32 ret = matSrc << 0; // Putting this as a separate statement fixes codegen, but regalloc is still wrong
-    return
-        ret |
-        enable << 1 |
-        (lightMask & 0x0F) << 2 |
-        ambSrc << 6 |
-        ((attnFn == GX_AF_SPEC) ? GX_DF_NONE : diffuseFn) << 7 |
-        (attnFn != GX_AF_NONE) << 9 |
-        (attnFn != GX_AF_SPEC) << 10 |
-        (lightMask >> 4 & 0x0F) << 11;
+static inline u32 setChanCtrlMacro(u8 enable, GXColorSrc ambSrc, GXColorSrc matSrc, u32 lightMask,
+                                   GXDiffuseFn diffuseFn, GXAttnFn attnFn) {
+    return matSrc << 0 | enable << 1 | (lightMask & 0x0F) << 2 | ambSrc << 6 |
+           ((attnFn == GX_AF_SPEC) ? GX_DF_NONE : diffuseFn) << 7 | (attnFn != GX_AF_NONE) << 9 |
+           (attnFn != GX_AF_SPEC) << 10 | (lightMask >> 4 & 0x0F) << 11;
 }
 
 struct J3DColorChan {
@@ -992,21 +986,22 @@ struct J3DColorChan {
             info.mDiffuseFn, info.mAttnFn, ambSrc);
     }
     J3DColorChan(u16 id) : mChanCtrl(id) {}
-    GXAttnFn getAttnFn() {
+    u8 getAttnFn() const {
         u8 attnFnTbl[] = { GX_AF_NONE, GX_AF_SPEC, GX_AF_NONE, GX_AF_SPOT };
-        return GXAttnFn(attnFnTbl[mChanCtrl >> 9 & 0x03]);
+        return attnFnTbl[(u32)(mChanCtrl & (3 << 9)) >> 9];
     }
-    GXDiffuseFn getDiffuseFn() { return GXDiffuseFn(mChanCtrl >> 7 & 3); }
-    u8 getLightMask() { return ((mChanCtrl >> 2 & 0x0f) | (mChanCtrl >> 11 & 0x0f) << 4); }
+    u8 getDiffuseFn() const { return ((u32)(mChanCtrl & (3 << 7)) >> 7); }
+    u8 getLightMask() const { return ((mChanCtrl >> 2) & 0x0f) | ((mChanCtrl >> 11) & 0x0f) << 4; }
     void setLightMask(u8 mask) {
         mChanCtrl = (mChanCtrl & ~0x003c) | ((mask & 0x0F) << 2);
         mChanCtrl = (mChanCtrl & ~0x7800) | ((mask & 0xF0) << 7);
     }
-    GXColorSrc getMatSrc() { return GXColorSrc(mChanCtrl >> 0 & 0x01); }
-    GXColorSrc getAmbSrc() { return GXColorSrc(mChanCtrl >> 6 & 0x01); }
-    u8 getEnable() { return !!(mChanCtrl & 0x02); }
-    void load() {
-        J3DGDWrite_u32(setChanCtrlMacro(getEnable(), getAmbSrc(), getMatSrc(), getLightMask(), getDiffuseFn(), getAttnFn()));
+    u8 getMatSrc() const { return (GXColorSrc)(mChanCtrl & 1); }
+    u8 getAmbSrc() const { return (GXColorSrc)((u32)(mChanCtrl & (1 << 6)) >> 6); }
+    u8 getEnable() const { return (u32)(mChanCtrl & 0x2) >> 1; }
+    void load() const {
+        J3DGDWrite_u32(setChanCtrlMacro(getEnable(), (GXColorSrc)getAmbSrc(), (GXColorSrc)getMatSrc(), getLightMask(),
+                                        (GXDiffuseFn)getDiffuseFn(), (GXAttnFn)getAttnFn()));
     }
 
     /* 0x0 */ u16 mChanCtrl;

src/JSystem/J3DGraphBase/J3DMatBlock.cpp

@@ -341,7 +341,6 @@ void J3DColorBlockLightOff::load() {
     mColorChan[2].load();
     mColorChan[1].load();
     mColorChan[3].load();
-    /* Nonmatching */
 }
 
 /* 802E0438-802E0AC0       .text load__22J3DColorBlockAmbientOnFv */
@@ -356,7 +355,6 @@ void J3DColorBlockAmbientOn::load() {
     mColorChan[2].load();
     mColorChan[1].load();
     mColorChan[3].load();
-    /* Nonmatching */
 }
 
 /* 802E0AC0-802E1180       .text load__20J3DColorBlockLightOnFv */
@@ -376,7 +374,6 @@ void J3DColorBlockLightOn::load() {
             mLight[i]->load(i);
         }
     }
-    /* Nonmatching */
 }
 
 /* 802E1180-802E11CC       .text patch__21J3DColorBlockLightOffFv */
@@ -407,7 +404,6 @@ void J3DColorBlockLightOff::patchLight() {
     mColorChan[3].load();
     u8* end = GDGetCurrPointer();
     DCFlushRange(start, end - start);
-    /* Nonmatching */
 }
 
 /* 802E17B4-802E1800       .text patch__20J3DColorBlockLightOnFv */
@@ -428,7 +424,6 @@ void J3DColorBlockLightOn::patchMatColor() {
 
 /* 802E19AC-802E1E18       .text patchLight__20J3DColorBlockLightOnFv */
 void J3DColorBlockLightOn::patchLight() {
-    /* Nonmatching */
     GDSetCurrOffset(mColorChanOffset);
     u8* start = GDGetCurrPointer();
     GDOverflowCheck(SizeOfLoadColorChans);
@@ -444,7 +439,6 @@ void J3DColorBlockLightOn::patchLight() {
     }
     u8* end = GDGetCurrPointer();
     DCFlushRange(start, end - start);
-    /* Nonmatching */
 }
 
 /* 802E1E18-802E1E80       .text diff__21J3DColorBlockLightOffFUl */
@@ -469,7 +463,6 @@ void J3DColorBlockLightOff::diffLight() {
     mColorChan[2].load();
     mColorChan[1].load();
     mColorChan[3].load();
-    /* Nonmatching */
 }
 
 /* 802E2408-802E2478       .text diff__20J3DColorBlockLightOnFUl */
@@ -1414,7 +1407,7 @@ void J3DTevBlock::indexToPtr_private(u32 offs) {
 
 /* 802E6298-802E6494       .text load__15J3DIndBlockFullFv */
 void J3DIndBlockFull::load() {
-    /* Nonmatching */
+    /* Nonmatching - regalloc */
     u8 indTexStageNum = mIndTexStageNum;
     for (u32 i = 0; i < indTexStageNum; i++) {
         mIndTexMtx[i].load(i);
@@ -1447,7 +1440,7 @@ void J3DIndBlockFull::load() {
 
 /* 802E6494-802E657C       .text diff__15J3DIndBlockFullFUl */
 void J3DIndBlockFull::diff(u32 flag) {
-    /* Nonmatching */
+    /* Nonmatching - regalloc */
     if ((flag & 0x08000000) == 0) {
         return;
     }
@@ -1455,7 +1448,7 @@ void J3DIndBlockFull::diff(u32 flag) {
     J3DGDSetIndTexStageNum(indTexStageNum);
     mIndTexMtx[0].load(0);
     J3DGDSetIndTexCoordScale(
-        GXIndTexStageID(0),
+        GX_INDTEXSTAGE0,
         GXIndTexScale(mIndTexCoordScale[0].getScaleS()),
         GXIndTexScale(mIndTexCoordScale[0].getScaleT()),
         GXIndTexScale(mIndTexCoordScale[1].getScaleS()),
@@ -1509,7 +1502,6 @@ void J3DPEBlockFogOff::load() {
     mBlend.load(mDither);
     mZMode.load();
     loadZCompLoc(mZCompLoc);
-    /* Nonmatching */
 }
 
 /* 802E7250-802E7538       .text diffBlend__16J3DPEBlockFogOffFv */
@@ -1517,7 +1509,6 @@ void J3DPEBlockFogOff::diffBlend() {
     GDOverflowCheck(0xf);
     mBlend.load(mDither);
     mZMode.load();
-    /* Nonmatching */
 }
 
 /* 802E7538-802E7A1C       .text load__14J3DPEBlockFullFv */
@@ -1531,7 +1522,6 @@ void J3DPEBlockFull::load() {
     mBlend.load(mDither);
     mZMode.load();
     loadZCompLoc(mZCompLoc);
-    /* Nonmatching */
 }
 
 /* 802E7A1C-802E7AD8       .text patch__14J3DPEBlockFullFv */
@@ -1544,7 +1534,6 @@ void J3DPEBlockFull::patch() {
     }
     u8* end = GDGetCurrPointer();
     DCFlushRange(start, end - start);
-    /* Nonmatching */
 }
 
 /* 802E7AD8-802E7B5C       .text diffFog__14J3DPEBlockFullFv */
@@ -1560,7 +1549,6 @@ void J3DPEBlockFull::diffBlend() {
     GDOverflowCheck(0xf);
     mBlend.load(mDither);
     mZMode.load();
-    /* Nonmatching */
 }
 
 /* 802E7E44-802E7EAC       .text diff__14J3DPEBlockFullFUl */