Merge pull request #4945 from Sonicadvance1/cachy_fexos

FEXCore: Adds option to disable L2 cache lookups

FEXCore/Source/Interface/Config/Config.json.in

@@ -161,6 +161,14 @@
         "Desc": [
           "Allows the user to pass additional arguments to the application"
         ]
+      },
+      "DisableL2Cache": {
+        "Type": "bool",
+        "Default": "false",
+        "Desc": [
+          "Disables FEXCore's JIT L2 cache lookup. Saving memory.",
+          "Can potentially introduce more stutters."
+        ]
       }
     },
     "Debug": {

FEXCore/Source/Interface/Core/Dispatcher/Dispatcher.cpp

@@ -177,60 +177,64 @@ void Dispatcher::EmitDispatcher() {
   ldrb(TMP1, STATE_PTR(CpuStateFrame, State.flags[X86State::RFLAG_TF_RAW_LOC]));
   cbnz(ARMEmitter::Size::i32Bit, TMP1, &CompileSingleStep);
 
-  // This is the block cache lookup routine
-  // It matches what is going on it LookupCache.h::FindBlock
-  ldr(TMP1, STATE_PTR(CpuStateFrame, Pointers.Common.L2Pointer));
-
-  // Mask the address by the virtual address size so we can check for aliases
-  uint64_t VirtualMemorySize = CTX->Config.VirtualMemSize;
-  if (std::popcount(VirtualMemorySize) == 1) {
-    and_(ARMEmitter::Size::i64Bit, TMP4, RipReg.R(), VirtualMemorySize - 1);
-  } else {
-    LoadConstant(ARMEmitter::Size::i64Bit, TMP4, VirtualMemorySize);
-    and_(ARMEmitter::Size::i64Bit, TMP4, RipReg.R(), TMP4);
-  }
-
   ARMEmitter::ForwardLabel NoBlock;
 
-  {
-    // Offset the address and add to our page pointer
-    lsr(ARMEmitter::Size::i64Bit, TMP2, TMP4, 12);
+  if (DisableL2Cache()) {
+    b(&NoBlock);
+  } else {
+    // This is the block cache lookup routine
+    // It matches what is going on it LookupCache.h::FindBlock
+    ldr(TMP1, STATE_PTR(CpuStateFrame, Pointers.Common.L2Pointer));
 
-    // Load the pointer from the offset
-    ldr(TMP1, TMP1, TMP2, ARMEmitter::ExtendedType::LSL_64, 3);
+    // Mask the address by the virtual address size so we can check for aliases
+    uint64_t VirtualMemorySize = CTX->Config.VirtualMemSize;
+    if (std::popcount(VirtualMemorySize) == 1) {
+      and_(ARMEmitter::Size::i64Bit, TMP4, RipReg.R(), VirtualMemorySize - 1);
+    } else {
+      LoadConstant(ARMEmitter::Size::i64Bit, TMP4, VirtualMemorySize);
+      and_(ARMEmitter::Size::i64Bit, TMP4, RipReg.R(), TMP4);
+    }
 
-    // If page pointer is zero then we have no block
-    cbz(ARMEmitter::Size::i64Bit, TMP1, &NoBlock);
-
-    // Steal the page offset
-    and_(ARMEmitter::Size::i64Bit, TMP2, TMP4, 0x0FFF);
-
-    // Shift the offset by the size of the block cache entry
-    add(TMP1, TMP1, TMP2, ARMEmitter::ShiftType::LSL, (int)log2(sizeof(FEXCore::LookupCache::LookupCacheEntry)));
-
-    // The the full LookupCacheEntry with a single LDP.
-    // Check the guest address first to ensure it maps to the address we are currently at.
-    // This fixes aliasing problems
-    ldp<ARMEmitter::IndexType::OFFSET>(TMP4, TMP2, TMP1, 0);
-
-    // If the guest address doesn't match, Compile the block.
-    sub(TMP2, TMP2, RipReg);
-    cbnz(ARMEmitter::Size::i64Bit, TMP2, &NoBlock);
-
-    // Check the host address to see if it matches, else compile the block.
-    cbz(ARMEmitter::Size::i64Bit, TMP4, &NoBlock);
-
-    // If we've made it here then we have a real compiled block
     {
-      // update L1 cache
-      ldr(TMP1, STATE_PTR(CpuStateFrame, Pointers.Common.L1Pointer));
+      // Offset the address and add to our page pointer
+      lsr(ARMEmitter::Size::i64Bit, TMP2, TMP4, 12);
 
-      and_(ARMEmitter::Size::i64Bit, TMP2, RipReg.R(), LookupCache::L1_ENTRIES_MASK);
-      add(TMP1, TMP1, TMP2, ARMEmitter::ShiftType::LSL, 4);
-      stp<ARMEmitter::IndexType::OFFSET>(TMP4, RipReg, TMP1);
+      // Load the pointer from the offset
+      ldr(TMP1, TMP1, TMP2, ARMEmitter::ExtendedType::LSL_64, 3);
 
-      // Jump to the block
-      br(TMP4);
+      // If page pointer is zero then we have no block
+      cbz(ARMEmitter::Size::i64Bit, TMP1, &NoBlock);
+
+      // Steal the page offset
+      and_(ARMEmitter::Size::i64Bit, TMP2, TMP4, 0x0FFF);
+
+      // Shift the offset by the size of the block cache entry
+      add(TMP1, TMP1, TMP2, ARMEmitter::ShiftType::LSL, (int)log2(sizeof(FEXCore::LookupCache::LookupCacheEntry)));
+
+      // The the full LookupCacheEntry with a single LDP.
+      // Check the guest address first to ensure it maps to the address we are currently at.
+      // This fixes aliasing problems
+      ldp<ARMEmitter::IndexType::OFFSET>(TMP4, TMP2, TMP1, 0);
+
+      // If the guest address doesn't match, Compile the block.
+      sub(TMP2, TMP2, RipReg);
+      cbnz(ARMEmitter::Size::i64Bit, TMP2, &NoBlock);
+
+      // Check the host address to see if it matches, else compile the block.
+      cbz(ARMEmitter::Size::i64Bit, TMP4, &NoBlock);
+
+      // If we've made it here then we have a real compiled block
+      {
+        // update L1 cache
+        ldr(TMP1, STATE_PTR(CpuStateFrame, Pointers.Common.L1Pointer));
+
+        and_(ARMEmitter::Size::i64Bit, TMP2, RipReg.R(), LookupCache::L1_ENTRIES_MASK);
+        add(TMP1, TMP1, TMP2, ARMEmitter::ShiftType::LSL, 4);
+        stp<ARMEmitter::IndexType::OFFSET>(TMP4, RipReg, TMP1);
+
+        // Jump to the block
+        br(TMP4);
+      }
     }
   }
 

FEXCore/Source/Interface/Core/Dispatcher/Dispatcher.h

@@ -4,6 +4,7 @@
 #include "Interface/Core/ArchHelpers/Arm64Emitter.h"
 #include "Interface/Core/Interpreter/InterpreterOps.h"
 
+#include <FEXCore/Config/Config.h>
 #include <FEXCore/fextl/memory.h>
 
 #include <array>
@@ -92,6 +93,8 @@ private:
 
   void EmitDispatcher();
   uint64_t GenerateABICall(FallbackABI ABI);
+
+  FEX_CONFIG_OPT(DisableL2Cache, DISABLEL2CACHE);
 };
 
 } // namespace FEXCore::CPU

FEXCore/Source/Interface/Core/LookupCache.h

@@ -141,22 +141,24 @@ public:
     // L2 and L3 need to be locked
     auto lk = Shared->AcquireWriteLock();
 
-    // Try L2
-    const auto PageIndex = (Address & (VirtualMemSize - 1)) >> 12;
-    const auto PageOffset = Address & (0x0FFF);
+    if (!DisableL2Cache()) {
+      // Try L2
+      const auto PageIndex = (Address & (VirtualMemSize - 1)) >> 12;
+      const auto PageOffset = Address & (0x0FFF);
 
-    const auto Pointers = reinterpret_cast<uintptr_t*>(PagePointer);
-    auto LocalPagePointer = Pointers[PageIndex];
+      const auto Pointers = reinterpret_cast<uintptr_t*>(PagePointer);
+      auto LocalPagePointer = Pointers[PageIndex];
 
-    // Do we a page pointer for this address?
-    if (LocalPagePointer) {
-      // Find there pointer for the address in the blocks
-      auto BlockPointers = reinterpret_cast<LookupCacheEntry*>(LocalPagePointer);
+      // Do we a page pointer for this address?
+      if (LocalPagePointer) {
+        // Find there pointer for the address in the blocks
+        auto BlockPointers = reinterpret_cast<LookupCacheEntry*>(LocalPagePointer);
 
-      if (BlockPointers[PageOffset].GuestCode == Address) {
-        L1Entry.GuestCode = Address;
-        L1Entry.HostCode = BlockPointers[PageOffset].HostCode;
-        return L1Entry.HostCode;
+        if (BlockPointers[PageOffset].GuestCode == Address) {
+          L1Entry.GuestCode = Address;
+          L1Entry.HostCode = BlockPointers[PageOffset].HostCode;
+          return L1Entry.HostCode;
+        }
       }
     }
 
@@ -209,22 +211,24 @@ public:
       // and it hasn't been thoroughly tested
     }
 
-    // Do full map
-    Address = Address & (VirtualMemSize - 1);
-    uint64_t PageOffset = Address & (0x0FFF);
-    Address >>= 12;
+    if (!DisableL2Cache()) {
+      // Do full map
+      Address = Address & (VirtualMemSize - 1);
+      uint64_t PageOffset = Address & (0x0FFF);
+      Address >>= 12;
 
-    uintptr_t* Pointers = reinterpret_cast<uintptr_t*>(PagePointer);
-    uint64_t LocalPagePointer = Pointers[Address];
-    if (!LocalPagePointer) {
-      // Page for this code didn't even exist, nothing to do
-      return ErasedAny;
+      uintptr_t* Pointers = reinterpret_cast<uintptr_t*>(PagePointer);
+      uint64_t LocalPagePointer = Pointers[Address];
+      if (!LocalPagePointer) {
+        // Page for this code didn't even exist, nothing to do
+        return ErasedAny;
+      }
+
+      // Page exists, just set the offset to zero
+      auto BlockPointers = reinterpret_cast<LookupCacheEntry*>(LocalPagePointer);
+      BlockPointers[PageOffset].GuestCode = 0;
+      BlockPointers[PageOffset].HostCode = 0;
     }
-
-    // Page exists, just set the offset to zero
-    auto BlockPointers = reinterpret_cast<LookupCacheEntry*>(LocalPagePointer);
-    BlockPointers[PageOffset].GuestCode = 0;
-    BlockPointers[PageOffset].HostCode = 0;
     return true;
   }
 
@@ -268,34 +272,37 @@ private:
     L1Entry.GuestCode = Address;
     L1Entry.HostCode = HostCode;
 
-    // Do ful map
-    auto FullAddress = Address;
-    Address = Address & (VirtualMemSize - 1);
+    if (!DisableL2Cache()) {
+      // Do ful map
+      auto FullAddress = Address;
+      Address = Address & (VirtualMemSize - 1);
 
-    uint64_t PageOffset = Address & (0x0FFF);
-    Address >>= 12;
-    uintptr_t* Pointers = reinterpret_cast<uintptr_t*>(PagePointer);
-    uint64_t LocalPagePointer = Pointers[Address];
-    if (!LocalPagePointer) {
-      // We don't have a page pointer for this address
-      // Allocate one now if we can
-      uintptr_t NewPageBacking = AllocateBackingForPage();
-      if (!NewPageBacking) {
-        // Couldn't allocate, clear L2 and retry
-        ClearL2Cache(lk);
-        CacheBlockMapping(Address, HostCode, lk);
-        return;
+      uint64_t PageOffset = Address & (0x0FFF);
+      Address >>= 12;
+
+      uintptr_t* Pointers = reinterpret_cast<uintptr_t*>(PagePointer);
+      uint64_t LocalPagePointer = Pointers[Address];
+      if (!LocalPagePointer) {
+        // We don't have a page pointer for this address
+        // Allocate one now if we can
+        uintptr_t NewPageBacking = AllocateBackingForPage();
+        if (!NewPageBacking) {
+          // Couldn't allocate, clear L2 and retry
+          ClearL2Cache(lk);
+          CacheBlockMapping(Address, HostCode, lk);
+          return;
+        }
+        Pointers[Address] = NewPageBacking;
+        LocalPagePointer = NewPageBacking;
       }
-      Pointers[Address] = NewPageBacking;
-      LocalPagePointer = NewPageBacking;
+
+      // Add the new pointer to the page block
+      auto BlockPointers = reinterpret_cast<LookupCacheEntry*>(LocalPagePointer);
+
+      // This silently replaces existing mappings
+      BlockPointers[PageOffset].GuestCode = FullAddress;
+      BlockPointers[PageOffset].HostCode = HostCode;
     }
-
-    // Add the new pointer to the page block
-    auto BlockPointers = reinterpret_cast<LookupCacheEntry*>(LocalPagePointer);
-
-    // This silently replaces existing mappings
-    BlockPointers[PageOffset].GuestCode = FullAddress;
-    BlockPointers[PageOffset].HostCode = HostCode;
   }
 
   uintptr_t AllocateBackingForPage() {
@@ -326,5 +333,7 @@ private:
 
   FEXCore::Context::ContextImpl* ctx;
   uint64_t VirtualMemSize {};
+
+  FEX_CONFIG_OPT(DisableL2Cache, DISABLEL2CACHE);
 };
 } // namespace FEXCore