fix ares64 tracer, pull in some more commits from upstream

waterbox/ares64/ares/ares/ares/node/debugger/tracer/instruction.hpp

@@ -28,9 +28,9 @@ struct Instruction : Tracer {
   }
 
   auto address(u64 address) -> bool {
-    address &= (1ull << _addressBits) - 1;  //mask upper bits of address
+    address &= (u128(1) << _addressBits) - 1;  //mask upper bits of address
     _address = address;
-    address >>= _addressMask;  //clip unneeded alignment bits (to reduce _masks size)
+    /*address >>= _addressMask;  //clip unneeded alignment bits (to reduce _masks size)
 
     if(_mask && updateMasks()) {
       if(_masks[address >> 3] & 1 << (address & 7)) return false;  //do not trace twice
@@ -48,7 +48,7 @@ struct Instruction : Tracer {
         _history[index] = _history[index + 1];
       }
       _history.last() = _address;
-    }
+    }*/
 
     return true;
   }
@@ -56,11 +56,11 @@ struct Instruction : Tracer {
   //mark an already-executed address as not executed yet for trace masking.
   //call when writing to executable RAM to support self-modifying code.
   auto invalidate(u64 address) -> void {
-    if(unlikely(_mask && updateMasks())) {
-      address &= (1ull << _addressBits) - 1;
+    /*if(unlikely(_mask && updateMasks())) {
+      address &= (u128(1) << _addressBits) - 1;
       address >>= _addressMask;
       _masks[address >> 3] &= ~(1 << (address & 7));
-    }
+    }*/
   }
 
   auto notify(const string& instruction, const string& context, const string& extra = {}) -> void {
@@ -72,7 +72,6 @@ struct Instruction : Tracer {
       );
       _omitted = 0;
     }
-
     string output{
       _component, "  ",
       hex(_address, _addressBits + 3 >> 2), "  ",

waterbox/ares64/ares/ares/n64/rsp/interpreter-vpu.cpp

@@ -314,11 +314,12 @@ auto RSP::SFV(cr128& vt, cr32& rs, s8 imm) -> void {
 template<u8 e>
 auto RSP::SHV(cr128& vt, cr32& rs, s8 imm) -> void {
   auto address = rs.u32 + imm * 16;
+  auto index = address & 7;
+  address &= ~7;
   for(u32 offset = 0; offset < 8; offset++) {
     auto byte = e + offset * 2;
     auto value = vt.byte(byte + 0 & 15) << 1 | vt.byte(byte + 1 & 15) >> 7;
-    dmem.write<Byte>(address, value);
-    address += 2;
+    dmem.write<Byte>(address + (index + offset * 2 & 15), value);
   }
 }
 
@@ -411,8 +412,8 @@ auto RSP::SWV(cr128& vt, cr32& rs, s8 imm) -> void {
   auto address = rs.u32 + imm * 16;
   auto start = e;
   auto end = start + 16;
-  auto base = address & 15;
-  address &= ~15;
+  auto base = address & 7;
+  address &= ~7;
   for(u32 offset = start; offset < end; offset++) {
     dmem.write<Byte>(address + (base++ & 15), vt.byte(offset & 15));
   }

waterbox/ares64/ares/ares/n64/rsp/recompiler.cpp

@@ -1,14 +1,14 @@
 auto RSP::Recompiler::pool() -> Pool* {
   if(context) return context;
 
-  u32 hashcode = 0;
+  nall::Hash::CRC32 hashcode;
   for(u32 offset : range(4096)) {
-    hashcode = (hashcode << 5) + hashcode + self.imem.read<Byte>(offset);
+    hashcode.input(self.imem.read<Byte>(offset));
   }
 
   PoolHashPair pair;
   pair.pool = (Pool*)allocator.acquire();
-  pair.hashcode = hashcode;
+  pair.hashcode = hashcode.value();
   if(auto result = pools.find(pair)) {
     return context = result->pool;
   }

waterbox/ares64/ares/ares/n64/vi/vi.cpp

@@ -90,7 +90,7 @@ auto VI::main() -> void {
       #if defined(VULKAN)
       }
       #endif
-	  #endif
+      #endif
 
       refreshed = true;
       screen->frame();