From ad1c540d6ed1b2888bb780ae6bd14941af13e648 Mon Sep 17 00:00:00 2001
From: Dario <dariosamo@gmail.com>
Date: Sun, 21 Jul 2024 17:18:35 -0300
Subject: [PATCH] Add unaligned load support for vertex PD, fix edge case in
 sync requirement detection.

---
 src/gbi/rt64_gbi.cpp   |  7 +++++--
 src/hle/rt64_rsp.cpp   | 12 ++++++++----
 src/hle/rt64_rsp.h     |  3 ++-
 src/hle/rt64_state.cpp |  1 -
 4 files changed, 15 insertions(+), 8 deletions(-)

diff --git a/src/gbi/rt64_gbi.cpp b/src/gbi/rt64_gbi.cpp
index fd6e175..b263a4b 100644
--- a/src/gbi/rt64_gbi.cpp
+++ b/src/gbi/rt64_gbi.cpp
@@ -44,6 +44,7 @@ namespace RT64 {
     //                  Constant                      Identifier                                   UCode                    LowP    NoN     ReJ     MVP     Point
     // 
     const GBIInstance   F3D_SM64                  = { "SW Version: 2.0D, 04-01-96 (SM64)",         GBIUCode::F3D,         { false,  false,  false,  false,  false } };
+    const GBIInstance   F3D_PD                    = { "SW Version: Unknown (PD)",                  GBIUCode::F3DPD,       { false,  false,  false,  false,  false } };
     const GBIInstance   F3DEX_1_21                = { "F3DEX 1.21",                                GBIUCode::F3DEX,       { false,  false,  false,  false,  false } };
     const GBIInstance   F3DLX_1_21_REJ            = { "F3DLX 1.21.Rej",                            GBIUCode::F3DEX,       { true,   false,  true,   false,  false } };
     const GBIInstance   F3DEX_1_23                = { "F3DEX 1.23",                                GBIUCode::F3DEX,       { false,  false,  false,  false,  false } };
@@ -71,8 +72,9 @@ namespace RT64 {
     // 
     //                  Length      Hash                    Known instances               
     //     
-    static std::array<GBISegment, 21> textSegments = {
+    static std::array<GBISegment, 22> textSegments = {
             GBISegment{ 0x1408,     0xF50165C013FCB8A2ULL,  { &F3D_SM64 } },
+            GBISegment{ 0x1418,     0xAEBF9966DD0486DDULL,  { &F3D_PD } },
             GBISegment{ 0x1430,     0x9A7772037D709388ULL,  { &F3DEX_1_21 } },
             GBISegment{ 0x13D0,     0x1BEA638E869B0195ULL,  { &F3DLX_1_21_REJ } }, // Needs confirmation.
             GBISegment{ 0x1430,     0xAC03DE5B7B1E710FULL,  { &F3DEX_1_23 } },
@@ -95,8 +97,9 @@ namespace RT64 {
             GBISegment{ 0x18C0,     0x9300F34F3B438634ULL,  { &S2DEX2_FIFO_2_08 } },
     };
 
-    static std::array<GBISegment, 21> dataSegments = {
+    static std::array<GBISegment, 22> dataSegments = {
             GBISegment{ 0x800,      0x276AC049785A7E70ULL,  { &F3D_SM64 } },
+            GBISegment{ 0x800,      0x72AD2373CEC74AA7ULL,  { &F3D_PD } },
             GBISegment{ 0x800,      0x4B5FDED20C137EC1ULL,  { &F3DEX_1_21 } },
             GBISegment{ 0x800,      0x3828B4F75B0A0E6AULL,  { &F3DEX_1_23 } },
             GBISegment{ 0x800,      0x484C6940F5072C39ULL,  { &F3DLX_1_21_REJ } }, // Needs confirmation.
diff --git a/src/hle/rt64_rsp.cpp b/src/hle/rt64_rsp.cpp
index 871e1c1..15eb4e8 100644
--- a/src/hle/rt64_rsp.cpp
+++ b/src/hle/rt64_rsp.cpp
@@ -91,11 +91,11 @@ namespace RT64 {
     }
 
     // Masks addresses as the RSP DMA hardware would.
-    uint32_t RSP::maskPhysicalAddress(uint32_t address) {
+    template<uint32_t mask> uint32_t RSP::maskPhysicalAddress(uint32_t address) {
         if (state->extended.extendRDRAM && ((address & 0xF0000000) == 0x80000000)) {
             return address - 0x80000000;
         }
-        return address & 0x00FFFFF8;
+        return address & mask;
     }
 
     // Performs a lookup in the segment table to convert the given address.
@@ -109,7 +109,11 @@ namespace RT64 {
     // Converts the given segmented address and then applies the RSP DMA physical address mask.
     // Used in cases where the RSP performs a DMA with a segmented address as the input. 
     uint32_t RSP::fromSegmentedMasked(uint32_t segAddress) {
-        return maskPhysicalAddress(fromSegmented(segAddress));
+        return maskPhysicalAddress<0x00FFFFF8>(fromSegmented(segAddress));
+    }
+
+    uint32_t RSP::fromSegmentedMaskedPD(uint32_t segAddress) {
+        return maskPhysicalAddress<0x00FFFFFC>(fromSegmented(segAddress));
     }
 
     void RSP::setSegment(uint32_t seg, uint32_t address) {
@@ -324,7 +328,7 @@ namespace RT64 {
             return;
         }
 
-        const uint32_t rdramAddress = fromSegmentedMasked(address);
+        const uint32_t rdramAddress = fromSegmentedMaskedPD(address);
         const VertexPD *dlVerts = reinterpret_cast<const VertexPD *>(state->fromRDRAM(rdramAddress));
         for (uint32_t i = 0; i < vtxCount; i++) {
             Vertex &dst = vertices[dstIndex + i];
diff --git a/src/hle/rt64_rsp.h b/src/hle/rt64_rsp.h
index f6563ab..ba8f7d8 100644
--- a/src/hle/rt64_rsp.h
+++ b/src/hle/rt64_rsp.h
@@ -221,9 +221,10 @@ namespace RT64 {
         void reset();
         Projection::Type getCurrentProjectionType() const;
         void addCurrentProjection(Projection::Type type);
-        uint32_t maskPhysicalAddress(uint32_t address);
+        template<uint32_t mask> uint32_t maskPhysicalAddress(uint32_t address);
         uint32_t fromSegmented(uint32_t segAddress);
         uint32_t fromSegmentedMasked(uint32_t segAddress);
+        uint32_t fromSegmentedMaskedPD(uint32_t segAddress);
         void setSegment(uint32_t seg, uint32_t address);
         void matrix(uint32_t address, uint8_t params);
         void popMatrix(uint32_t count);
diff --git a/src/hle/rt64_state.cpp b/src/hle/rt64_state.cpp
index 2b15096..5d8d3cc 100644
--- a/src/hle/rt64_state.cpp
+++ b/src/hle/rt64_state.cpp
@@ -975,7 +975,6 @@ namespace RT64 {
                             const DrawCallTile &callTile = callTiles[callDesc.tileIndex + t];
                             if (callTile.rawTMEM) {
                                 flags.canDecodeTMEM = true;
-                                break;
                             }
 
                             if (callTile.syncRequired && !callTile.tileCopyUsed) {