Add DolphinCPalette, logic matched, unfortunately has instruction swaps around TLUT ops

Former-commit-id: 9684ace9af798ebc96dc7ef72537a60c5c071a6d

.vscode/settings.json

@@ -40,7 +40,8 @@
     "oscontext.h": "c",
     "oscache.h": "c",
     "stdlib.h": "c",
-    "musyx_priv.h": "c"
+    "musyx_priv.h": "c",
+    "gxenum.h": "c"
   },
   "files.autoSave": "onFocusChange",
   "files.insertFinalNewline": true,

asm/Kyoto/CFrameDelayedKiller.s

@@ -167,8 +167,8 @@ lbl_8036CBF8:
 /* 8036CC14 00369B74  38 21 00 20 */	addi r1, r1, 0x20
 /* 8036CC18 00369B78  4E 80 00 20 */	blr
 
-.global sub_8036cc1c
+.global sub_8036cc1c__19CFrameDelayedKillerFbPv
-sub_8036cc1c:
+sub_8036cc1c__19CFrameDelayedKillerFbPv:
 /* 8036CC1C 00369B7C  94 21 FF E0 */	stwu r1, -0x20(r1)
 /* 8036CC20 00369B80  7C 08 02 A6 */	mflr r0
 /* 8036CC24 00369B84  2C 03 00 01 */	cmpwi r3, 1

asm/Kyoto/Graphics/DolphinCModel.s

@@ -902,7 +902,7 @@ __dt__6CModelFv:
 /* 80355C98 00352BF8  38 00 00 00 */	li r0, 0
 /* 80355C9C 00352BFC  38 60 00 01 */	li r3, 1
 /* 80355CA0 00352C00  90 1E 00 00 */	stw r0, 0(r30)
-/* 80355CA4 00352C04  48 01 6F 79 */	bl sub_8036cc1c
+/* 80355CA4 00352C04  48 01 6F 79 */	bl sub_8036cc1c__19CFrameDelayedKillerFbPv
 /* 80355CA8 00352C08  48 00 00 24 */	b lbl_80355CCC
 lbl_80355CAC:
 /* 80355CAC 00352C0C  38 03 FF FF */	addi r0, r3, -1
@@ -912,7 +912,7 @@ lbl_80355CAC:
 /* 80355CBC 00352C1C  38 00 00 00 */	li r0, 0
 /* 80355CC0 00352C20  38 60 00 00 */	li r3, 0
 /* 80355CC4 00352C24  90 1E 00 00 */	stw r0, 0(r30)
-/* 80355CC8 00352C28  48 01 6F 55 */	bl sub_8036cc1c
+/* 80355CC8 00352C28  48 01 6F 55 */	bl sub_8036cc1c__19CFrameDelayedKillerFbPv
 lbl_80355CCC:
 /* 80355CCC 00352C2C  34 1E 00 28 */	addic. r0, r30, 0x28
 /* 80355CD0 00352C30  41 82 00 0C */	beq lbl_80355CDC
@@ -1523,4 +1523,3 @@ lbl_803D8308:
 	# ROM: 0x3D5308
 	.asciz "??(??)"
 	.balign 4
-

asm/Kyoto/Graphics/DolphinCPalette.s

@@ -3,8 +3,8 @@
 .section .sbss, "wa"
 .balign 8
 
-.global lbl_805A9428
+.global sCurrentFrameCount__16CGraphicsPalette
-lbl_805A9428:
+sCurrentFrameCount__16CGraphicsPalette:
 	.skip 0x8
 
 .section .text, "ax"
@@ -48,7 +48,7 @@ Load__16CGraphicsPaletteCFv:
 /* 8030DED8 0030AE38  7C 7F 1B 78 */	mr r31, r3
 /* 8030DEDC 0030AE3C  38 7F 00 10 */	addi r3, r31, 0x10
 /* 8030DEE0 0030AE40  48 06 C6 51 */	bl GXLoadTlut
-/* 8030DEE4 0030AE44  80 0D A8 68 */	lwz r0, lbl_805A9428@sda21(r13)
+/* 8030DEE4 0030AE44  80 0D A8 68 */	lwz r0, sCurrentFrameCount__16CGraphicsPalette@sda21(r13)
 /* 8030DEE8 0030AE48  90 1F 00 04 */	stw r0, 4(r31)
 /* 8030DEEC 0030AE4C  80 01 00 14 */	lwz r0, 0x14(r1)
 /* 8030DEF0 0030AE50  83 E1 00 0C */	lwz r31, 0xc(r1)
@@ -67,7 +67,7 @@ __dt__16CGraphicsPaletteFv:
 /* 8030DF18 0030AE78  7C 7E 1B 79 */	or. r30, r3, r3
 /* 8030DF1C 0030AE7C  41 82 00 50 */	beq lbl_8030DF6C
 /* 8030DF20 0030AE80  80 7E 00 04 */	lwz r3, 4(r30)
-/* 8030DF24 0030AE84  80 0D A8 68 */	lwz r0, lbl_805A9428@sda21(r13)
+/* 8030DF24 0030AE84  80 0D A8 68 */	lwz r0, sCurrentFrameCount__16CGraphicsPalette@sda21(r13)
 /* 8030DF28 0030AE88  7C 03 00 50 */	subf r0, r3, r0
 /* 8030DF2C 0030AE8C  28 00 00 02 */	cmplwi r0, 2
 /* 8030DF30 0030AE90  40 80 00 1C */	bge lbl_8030DF4C
@@ -76,7 +76,7 @@ __dt__16CGraphicsPaletteFv:
 /* 8030DF3C 0030AE9C  7C 00 00 34 */	cntlzw r0, r0
 /* 8030DF40 0030AEA0  90 7E 00 0C */	stw r3, 0xc(r30)
 /* 8030DF44 0030AEA4  54 03 D9 7E */	srwi r3, r0, 5
-/* 8030DF48 0030AEA8  48 05 EC D5 */	bl sub_8036cc1c
+/* 8030DF48 0030AEA8  48 05 EC D5 */	bl sub_8036cc1c__19CFrameDelayedKillerFbPv
 lbl_8030DF4C:
 /* 8030DF4C 0030AEAC  34 1E 00 0C */	addic. r0, r30, 0xc
 /* 8030DF50 0030AEB0  41 82 00 0C */	beq lbl_8030DF5C

asm/Kyoto/Graphics/DolphinCTexture.s

@@ -66,7 +66,7 @@ sub_8030e10c__8CTextureFv:
 /* 8030E128 0030B088  48 04 A2 AD */	bl sub_803583d4
 /* 8030E12C 0030B08C  7C 64 1B 78 */	mr r4, r3
 /* 8030E130 0030B090  38 60 00 01 */	li r3, 1
-/* 8030E134 0030B094  48 05 EA E9 */	bl sub_8036cc1c
+/* 8030E134 0030B094  48 05 EA E9 */	bl sub_8036cc1c__19CFrameDelayedKillerFbPv
 lbl_8030E138:
 /* 8030E138 0030B098  80 01 00 14 */	lwz r0, 0x14(r1)
 /* 8030E13C 0030B09C  7C 08 03 A6 */	mtlr r0

asm/MetroidPrime/CMapArea.s

@@ -1007,7 +1007,7 @@ __dt__8CMapAreaFv:
 /* 80080EB0 0007DE10  90 8D A2 F0 */	stw r4, lbl_805A8EB0@sda21(r13)
 /* 80080EB4 0007DE14  80 9E 00 44 */	lwz r4, 0x44(r30)
 /* 80080EB8 0007DE18  90 1E 00 44 */	stw r0, 0x44(r30)
-/* 80080EBC 0007DE1C  48 2E BD 61 */	bl sub_8036cc1c
+/* 80080EBC 0007DE1C  48 2E BD 61 */	bl sub_8036cc1c__19CFrameDelayedKillerFbPv
 /* 80080EC0 0007DE20  34 1E 00 44 */	addic. r0, r30, 0x44
 /* 80080EC4 0007DE24  41 82 00 0C */	beq lbl_80080ED0
 /* 80080EC8 0007DE28  80 7E 00 44 */	lwz r3, 0x44(r30)
@@ -1146,4 +1146,3 @@ lbl_803CD66C:
 	.asciz "??(??)"
 	.balign 4
 	.4byte 0
-

asm/MetroidPrime/CStateManager.s

@@ -8991,7 +8991,7 @@ FrameBegin__13CStateManagerFUi:
 /* 8004B574 000484D4  80 03 08 D4 */	lwz r0, 0x8d4(r3)
 /* 8004B578 000484D8  90 0D A8 70 */	stw r0, lbl_805A9430@sda21(r13)
 /* 8004B57C 000484DC  80 03 08 D4 */	lwz r0, 0x8d4(r3)
-/* 8004B580 000484E0  90 0D A8 68 */	stw r0, lbl_805A9428@sda21(r13)
+/* 8004B580 000484E0  90 0D A8 68 */	stw r0, sCurrentFrameCount__16CGraphicsPalette@sda21(r13)
 /* 8004B584 000484E4  4B FF FF D5 */	bl SwapOutTexturesToARAM__13CStateManagerFiUi
 /* 8004B588 000484E8  80 01 00 14 */	lwz r0, 0x14(r1)
 /* 8004B58C 000484EC  7C 08 03 A6 */	mtlr r0

configure.py

@@ -558,8 +558,8 @@ LIBS = [
             "Kyoto/Graphics/CCubeModel",
             ["Kyoto/Graphics/CGX", True],
             ["Kyoto/Graphics/CTevCombiners", True],
-            "Kyoto/Graphics/DolphinCGraphics",
+            ["Kyoto/Graphics/DolphinCGraphics", False],
-            "Kyoto/Graphics/DolphinCPalette",
+            ["Kyoto/Graphics/DolphinCPalette", False],
             "Kyoto/Graphics/DolphinCTexture",
             ["Kyoto/Math/CloseEnough", True],
             "Kyoto/Math/CMatrix3f",

include/Kyoto/CFrameDelayedKiller.hpp

@@ -9,6 +9,7 @@ public:
   static void ShutDown();
 
   static void sub_8036cb90();
+  static void sub_8036cc1c(bool unk1, void* unk2);
 };
 
 #endif // _CFRAMEDELAYEDKILLER

include/Kyoto/Graphics/CGraphicsPalette.hpp

@@ -0,0 +1,39 @@
+#ifndef _CGRAPHICSPALETTE
+#define _CGRAPHICSPALETTE
+
+#include "types.h"
+
+#include "dolphin/gx.h"
+
+#include <rstl/single_ptr.hpp>
+
+enum EPaletteFormat {
+  kPF_IA8 = GX_TL_IA8,
+  kPF_RGB565 = GX_TL_RGB565,
+  kPF_RGB5A3 = GX_TL_RGB5A3,
+};
+
+class CInputStream;
+
+class CGraphicsPalette {
+public:
+  CGraphicsPalette(EPaletteFormat format, int numEntries);
+  CGraphicsPalette(CInputStream& in);
+  ~CGraphicsPalette();
+
+  ushort* GetPaletteData() { return xc_entries.get(); }
+  const ushort* GetPaletteData() const { return xc_entries.get(); }
+  void Load() const;
+  void UnLock();
+
+private:
+  static uint sCurrentFrameCount;
+  EPaletteFormat x0_fmt;
+  mutable uint x4_frameLoaded;
+  uint x8_entryCount;
+  rstl::single_ptr<ushort> xc_entries;
+  GXTlutObj x10_tlutObj;
+  bool x1c_locked;
+};
+
+#endif // _CGRAPHICSPALETTE

src/Kyoto/Graphics/DolphinCPalette.cpp

@@ -0,0 +1,45 @@
+#include "Kyoto/Graphics/CGraphicsPalette.hpp"
+
+#include "Kyoto/Alloc/CMemorySys.hpp"
+#include "Kyoto/CFrameDelayedKiller.hpp"
+#include "Kyoto/Streams/CInputStream.hpp"
+#include "dolphin/os/OSCache.h"
+
+uint CGraphicsPalette::sCurrentFrameCount = 0;
+
+CGraphicsPalette::CGraphicsPalette(EPaletteFormat format, int numEntries)
+: x0_fmt(format)
+, x8_entryCount(numEntries)
+, xc_entries((ushort*)CMemory::Alloc(numEntries * sizeof(ushort), IAllocator::kHI_RoundUpLen))
+, x1c_locked(false) {
+  GXInitTlutObj(&x10_tlutObj, xc_entries.get(), (GXTlutFmt)(x0_fmt), x8_entryCount);
+}
+
+CGraphicsPalette::CGraphicsPalette(CInputStream& in)
+: x0_fmt(EPaletteFormat(in.ReadLong()))
+, x8_entryCount(in.Get< short >() * in.Get< short >())
+, xc_entries((ushort*)CMemory::Alloc(x8_entryCount * sizeof(ushort), IAllocator::kHI_RoundUpLen))
+, x1c_locked(false) {
+  in.Get(reinterpret_cast< uchar* >(xc_entries.get()), x8_entryCount * sizeof(ushort));
+  GXInitTlutObj(&x10_tlutObj, xc_entries.get(), static_cast< GXTlutFmt >(x0_fmt), x8_entryCount);
+  DCFlushRange(xc_entries.get(), x8_entryCount * sizeof(ushort));
+}
+
+CGraphicsPalette::~CGraphicsPalette() {
+  uint frameDiff = sCurrentFrameCount - x4_frameLoaded;
+  if (frameDiff < 2) {
+    CFrameDelayedKiller::sub_8036cc1c(frameDiff == 0, xc_entries.release());
+  }
+}
+
+void CGraphicsPalette::Load() const {
+  GXLoadTlut(&x10_tlutObj, GX_TLUT0);
+  x4_frameLoaded = sCurrentFrameCount;
+}
+
+void CGraphicsPalette::UnLock() {
+  DCStoreRange(xc_entries.get(), x8_entryCount * sizeof(ushort));
+  GXInitTlutObj(&x10_tlutObj, xc_entries.get(), static_cast< GXTlutFmt >(x0_fmt), x8_entryCount);
+  DCFlushRange(xc_entries.get(), x8_entryCount * sizeof(ushort));
+  x1c_locked = false;
+}