Merge pull request #8731 from Pokechu22/dsp-lle-cpu-int-crash

Fix crash when using DSP LLE with CPU interpreter (or fastmem off)

Source/Core/Common/Hash.cpp

@@ -95,7 +95,7 @@ u32 HashAdler32(const u8* data, size_t len)
 
 // Stupid hash - but can't go back now :)
 // Don't use for new things. At least it's reasonably fast.
-u32 HashEctor(const u8* ptr, int length)
+u32 HashEctor(const u8* ptr, size_t length)
 {
   u32 crc = 0;
 

Source/Core/Common/Hash.h

@@ -12,7 +12,7 @@ namespace Common
 {
 u32 HashFletcher(const u8* data_u8, size_t length);  // FAST. Length & 1 == 0.
 u32 HashAdler32(const u8* data, size_t len);         // Fairly accurate, slightly slower
-u32 HashEctor(const u8* ptr, int length);            // JUNK. DO NOT USE FOR NEW THINGS
+u32 HashEctor(const u8* ptr, size_t length);         // JUNK. DO NOT USE FOR NEW THINGS
 u64 GetHash64(const u8* src, u32 len, u32 samples);
 void SetHash64Function();
 }  // namespace Common

Source/Core/Core/DSP/DSPCodeUtil.cpp

@@ -68,11 +68,11 @@ bool Compare(const std::vector<u16>& code1, const std::vector<u16>& code2)
   if (code1.size() != code2.size())
     printf("Size difference! 1=%zu 2=%zu\n", code1.size(), code2.size());
   u32 count_equal = 0;
-  const int min_size = std::min<int>((int)code1.size(), (int)code2.size());
+  const u16 min_size = static_cast<u16>(std::min(code1.size(), code2.size()));
 
   AssemblerSettings settings;
   DSPDisassembler disassembler(settings);
-  for (int i = 0; i < min_size; i++)
+  for (u16 i = 0; i < min_size; i++)
   {
     if (code1[i] == code2[i])
     {
@@ -93,7 +93,7 @@ bool Compare(const std::vector<u16>& code1, const std::vector<u16>& code2)
   {
     printf("Extra code words:\n");
     const std::vector<u16>& longest = code1.size() > code2.size() ? code1 : code2;
-    for (int i = min_size; i < (int)longest.size(); i++)
+    for (u16 i = min_size; i < longest.size(); i++)
     {
       u16 pc = i;
       std::string line;
@@ -146,7 +146,7 @@ bool SaveBinary(const std::vector<u16>& code, const std::string& filename)
   return File::WriteStringToFile(filename, buffer);
 }
 
-bool DumpDSPCode(const u8* code_be, int size_in_bytes, u32 crc)
+bool DumpDSPCode(const u8* code_be, size_t size_in_bytes, u32 crc)
 {
   const std::string root_name =
       File::GetUserPath(D_DUMPDSP_IDX) + fmt::format("DSP_UC_{:08X}", crc);

Source/Core/Core/DSP/DSPCodeUtil.h

@@ -24,5 +24,5 @@ std::vector<u16> BinaryStringBEToCode(const std::string& str);
 std::optional<std::vector<u16>> LoadBinary(const std::string& filename);
 bool SaveBinary(const std::vector<u16>& code, const std::string& filename);
 
-bool DumpDSPCode(const u8* code_be, int size_in_bytes, u32 crc);
+bool DumpDSPCode(const u8* code_be, size_t size_in_bytes, u32 crc);
 }  // namespace DSP

Source/Core/Core/DSP/DSPCore.h

@@ -306,9 +306,6 @@ struct SDSP
   u16* dram;
   u16* irom;
   u16* coef;
-
-  // This one doesn't really belong here.
-  u8* cpu_ram;
 };
 
 extern SDSP g_dsp;

Source/Core/Core/DSP/DSPHWInterface.cpp

@@ -11,7 +11,6 @@
 
 #include "Common/CPUDetect.h"
 #include "Common/CommonTypes.h"
-#include "Common/Hash.h"
 #include "Common/Intrinsics.h"
 #include "Common/Logging/Log.h"
 #include "Common/MemoryUtil.h"
@@ -283,22 +282,15 @@ u16 gdsp_ifx_read(u16 addr)
 
 static const u8* gdsp_idma_in(u16 dsp_addr, u32 addr, u32 size)
 {
-  u16* dst = g_dsp.iram + (dsp_addr / 2);
-
-  const u8* code = &g_dsp.cpu_ram[addr & 0x0fffffff];
-  g_dsp.iram_crc = Common::HashEctor(code, size);
-
   Common::UnWriteProtectMemory(g_dsp.iram, DSP_IRAM_BYTE_SIZE, false);
-  memcpy(dst, code, size);
-  for (size_t i = 0; i < size / 2; i++)
-    dst[i] = Common::swap16(dst[i]);
+  Host::DMAToDSP(g_dsp.iram + dsp_addr / 2, addr, size);
   Common::WriteProtectMemory(g_dsp.iram, DSP_IRAM_BYTE_SIZE, false);
 
-  Host::CodeLoaded(code, size);
+  Host::CodeLoaded(addr, size);
   NOTICE_LOG(DSPLLE, "*** Copy new UCode from 0x%08x to 0x%04x (crc: %8x)", addr, dsp_addr,
              g_dsp.iram_crc);
 
-  return reinterpret_cast<u8*>(dst);
+  return reinterpret_cast<u8*>(g_dsp.iram) + dsp_addr;
 }
 
 static const u8* gdsp_idma_out(u16 dsp_addr, u32 addr, u32 size)
@@ -309,80 +301,23 @@ static const u8* gdsp_idma_out(u16 dsp_addr, u32 addr, u32 size)
   return nullptr;
 }
 
-#if defined(_M_X86) || defined(_M_X86_64)
-static const __m128i s_mask = _mm_set_epi32(0x0E0F0C0DL, 0x0A0B0809L, 0x06070405L, 0x02030001L);
-
-FUNCTION_TARGET_SSSE3
-static void gdsp_ddma_in_SSSE3(u16 dsp_addr, u32 addr, u32 size, u8* dst)
-{
-  for (u32 i = 0; i < size; i += 16)
-  {
-    _mm_storeu_si128(
-        (__m128i*)&dst[dsp_addr + i],
-        _mm_shuffle_epi8(_mm_loadu_si128((__m128i*)&g_dsp.cpu_ram[(addr + i) & 0x7FFFFFFF]),
-                         s_mask));
-  }
-}
-
-FUNCTION_TARGET_SSSE3
-static void gdsp_ddma_out_SSSE3(u16 dsp_addr, u32 addr, u32 size, const u8* src)
-{
-  for (u32 i = 0; i < size; i += 16)
-  {
-    _mm_storeu_si128((__m128i*)&g_dsp.cpu_ram[(addr + i) & 0x7FFFFFFF],
-                     _mm_shuffle_epi8(_mm_loadu_si128((__m128i*)&src[dsp_addr + i]), s_mask));
-  }
-}
-#endif
-
 // TODO: These should eat clock cycles.
 static const u8* gdsp_ddma_in(u16 dsp_addr, u32 addr, u32 size)
 {
-  u8* dst = reinterpret_cast<u8*>(g_dsp.dram);
-
-#if defined(_M_X86) || defined(_M_X86_64)
-  if (cpu_info.bSSSE3 && !(size % 16))
-  {
-    gdsp_ddma_in_SSSE3(dsp_addr, addr, size, dst);
-  }
-  else
-#endif
-  {
-    for (u32 i = 0; i < size; i += 2)
-    {
-      *(u16*)&dst[dsp_addr + i] =
-          Common::swap16(*(const u16*)&g_dsp.cpu_ram[(addr + i) & 0x7FFFFFFF]);
-    }
-  }
+  Host::DMAToDSP(g_dsp.dram + dsp_addr / 2, addr, size);
   DEBUG_LOG(DSPLLE, "*** ddma_in RAM (0x%08x) -> DRAM_DSP (0x%04x) : size (0x%08x)", addr,
             dsp_addr / 2, size);
 
-  return dst + dsp_addr;
+  return reinterpret_cast<u8*>(g_dsp.dram) + dsp_addr;
 }
 
 static const u8* gdsp_ddma_out(u16 dsp_addr, u32 addr, u32 size)
 {
-  const u8* src = reinterpret_cast<const u8*>(g_dsp.dram);
-
-#ifdef _M_X86
-  if (cpu_info.bSSSE3 && !(size % 16))
-  {
-    gdsp_ddma_out_SSSE3(dsp_addr, addr, size, src);
-  }
-  else
-#endif
-  {
-    for (u32 i = 0; i < size; i += 2)
-    {
-      *(u16*)&g_dsp.cpu_ram[(addr + i) & 0x7FFFFFFF] =
-          Common::swap16(*(const u16*)&src[dsp_addr + i]);
-    }
-  }
-
+  Host::DMAFromDSP(g_dsp.dram + dsp_addr / 2, addr, size);
   DEBUG_LOG(DSPLLE, "*** ddma_out DRAM_DSP (0x%04x) -> RAM (0x%08x) : size (0x%08x)", dsp_addr / 2,
             addr, size);
 
-  return src + dsp_addr;
+  return reinterpret_cast<const u8*>(g_dsp.dram) + dsp_addr;
 }
 
 static void gdsp_do_dma()

Source/Core/Core/DSP/DSPHost.h

@@ -17,10 +17,13 @@ namespace DSP::Host
 {
 u8 ReadHostMemory(u32 addr);
 void WriteHostMemory(u8 value, u32 addr);
+void DMAToDSP(u16* dst, u32 addr, u32 size);
+void DMAFromDSP(const u16* src, u32 addr, u32 size);
 void OSD_AddMessage(std::string str, u32 ms);
 bool OnThread();
 bool IsWiiHost();
 void InterruptRequest();
-void CodeLoaded(const u8* ptr, int size);
+void CodeLoaded(u32 addr, size_t size);
+void CodeLoaded(const u8* ptr, size_t size);
 void UpdateDebugger();
 }  // namespace DSP::Host

Source/Core/Core/HW/DSPLLE/DSPHost.cpp

@@ -16,6 +16,7 @@
 #include "Core/DSP/Jit/x64/DSPEmitter.h"
 #include "Core/HW/DSP.h"
 #include "Core/HW/DSPLLE/DSPSymbols.h"
+#include "Core/HW/Memmap.h"
 #include "Core/Host.h"
 #include "VideoCommon/OnScreenDisplay.h"
 
@@ -36,6 +37,16 @@ void WriteHostMemory(u8 value, u32 addr)
   DSP::WriteARAM(value, addr);
 }
 
+void DMAToDSP(u16* dst, u32 addr, u32 size)
+{
+  Memory::CopyFromEmuSwapped(dst, addr, size);
+}
+
+void DMAFromDSP(const u16* src, u32 addr, u32 size)
+{
+  Memory::CopyToEmuSwapped(addr, src, size);
+}
+
 void OSD_AddMessage(std::string str, u32 ms)
 {
   OSD::AddMessage(std::move(str), ms);
@@ -57,8 +68,14 @@ void InterruptRequest()
   DSP::GenerateDSPInterruptFromDSPEmu(DSP::INT_DSP);
 }
 
-void CodeLoaded(const u8* ptr, int size)
+void CodeLoaded(u32 addr, size_t size)
 {
+  CodeLoaded(Memory::GetPointer(addr), size);
+}
+
+void CodeLoaded(const u8* ptr, size_t size)
+{
+  g_dsp.iram_crc = Common::HashEctor(ptr, size);
   if (SConfig::GetInstance().m_DumpUCode)
   {
     DSP::DumpDSPCode(ptr, size, g_dsp.iram_crc);

Source/Core/Core/HW/DSPLLE/DSPLLE.cpp

@@ -78,8 +78,10 @@ void DSPLLE::DoState(PointerWrap& p)
   Common::UnWriteProtectMemory(g_dsp.iram, DSP_IRAM_BYTE_SIZE, false);
   p.DoArray(g_dsp.iram, DSP_IRAM_SIZE);
   Common::WriteProtectMemory(g_dsp.iram, DSP_IRAM_BYTE_SIZE, false);
+  // TODO: This uses the wrong endianness (producing bad disassembly)
+  // and a bogus byte count (producing bad hashes)
   if (p.GetMode() == PointerWrap::MODE_READ)
-    Host::CodeLoaded((const u8*)g_dsp.iram, DSP_IRAM_BYTE_SIZE);
+    Host::CodeLoaded(reinterpret_cast<const u8*>(g_dsp.iram), DSP_IRAM_BYTE_SIZE);
   p.DoArray(g_dsp.dram, DSP_DRAM_SIZE);
   p.Do(g_init_hax);
   p.Do(m_cycle_count);
@@ -186,10 +188,6 @@ bool DSPLLE::Initialize(bool wii, bool dsp_thread)
   m_wii = wii;
   m_is_dsp_on_thread = dsp_thread;
 
-  // DSPLLE directly accesses the fastmem arena.
-  // TODO: The fastmem arena is only supposed to be used by the JIT:
-  // among other issues, its size is only 1GB on 32-bit targets.
-  g_dsp.cpu_ram = Memory::physical_base;
   DSPCore_Reset();
 
   InitInstructionTable();

Source/DSPTool/DSPTool.cpp

@@ -22,6 +22,12 @@ u8 DSP::Host::ReadHostMemory(u32 addr)
 void DSP::Host::WriteHostMemory(u8 value, u32 addr)
 {
 }
+void DSP::Host::DMAToDSP(u16* dst, u32 addr, u32 size)
+{
+}
+void DSP::Host::DMAFromDSP(const u16* src, u32 addr, u32 size)
+{
+}
 void DSP::Host::OSD_AddMessage(std::string str, u32 ms)
 {
 }
@@ -33,7 +39,10 @@ bool DSP::Host::IsWiiHost()
 {
   return false;
 }
-void DSP::Host::CodeLoaded(const u8* ptr, int size)
+void DSP::Host::CodeLoaded(u32 addr, size_t size)
+{
+}
+void DSP::Host::CodeLoaded(const u8* ptr, size_t size)
 {
 }
 void DSP::Host::InterruptRequest()