Merge pull request #2994 from Sonicadvance1/cache_namedvectorconstants

OpcodeDispatcher: Cache named vector constants in the block

FEXCore/Source/Interface/Core/OpcodeDispatcher.h

@@ -12,6 +12,7 @@
 #include <FEXCore/IR/IREmitter.h>
 
 #include <FEXCore/Utils/LogManager.h>
+#include <FEXCore/Utils/MathUtils.h>
 #include <FEXCore/fextl/map.h>
 #include <FEXCore/fextl/vector.h>
 
@@ -90,6 +91,9 @@ public:
 
     // New block needs to reset segment telemetry.
     SegmentsNeedReadCheck = ~0U;
+
+    // Need to clear any named constants that were cached.
+    ClearCachedNamedConstants();
   }
 
   bool FinishOp(uint64_t NextRIP, bool LastOp) {
@@ -1196,6 +1200,28 @@ private:
     }
   }
 
+  // Named constant cache for the current block.
+  // Different arrays for sizes 1,2,4,8,16,32.
+  OrderedNode *CachedNamedVectorConstants[FEXCore::IR::NamedVectorConstant::NAMED_VECTOR_MAX][6]{};
+
+  // Load and cache a named vector constant.
+  OrderedNode *LoadAndCacheNamedVectorConstant(uint8_t Size, FEXCore::IR::NamedVectorConstant NamedConstant) {
+    auto log2_size_bytes = FEXCore::ilog2(Size);
+    if (CachedNamedVectorConstants[NamedConstant][log2_size_bytes]) {
+      return CachedNamedVectorConstants[NamedConstant][log2_size_bytes];
+    }
+
+    auto Constant = _LoadNamedVectorConstant(Size, NamedConstant);
+    CachedNamedVectorConstants[NamedConstant][log2_size_bytes] = Constant;
+    return Constant;
+  }
+
+  // Reset the named vector constants cache array.
+  // These are only cached per block.
+  void ClearCachedNamedConstants() {
+    memset(CachedNamedVectorConstants, 0, sizeof(CachedNamedVectorConstants));
+  }
+
   OrderedNode *SelectCC(uint8_t OP, OrderedNode *TrueValue, OrderedNode *FalseValue);
 
   /**

FEXCore/Source/Interface/Core/OpcodeDispatcher/Vector.cpp

@@ -2962,7 +2962,7 @@ OrderedNode* OpDispatchBuilder::ADDSUBPOpImpl(OpcodeArgs, size_t ElementSize,
                                               OrderedNode *Src1, OrderedNode *Src2) {
   const auto Size = GetSrcSize(Op);
 
-  auto ConstantEOR = _LoadNamedVectorConstant(Size, ElementSize == 4 ? NAMED_VECTOR_PADDSUBPS_INVERT : NAMED_VECTOR_PADDSUBPD_INVERT);
+  auto ConstantEOR = LoadAndCacheNamedVectorConstant(Size, ElementSize == 4 ? NAMED_VECTOR_PADDSUBPS_INVERT : NAMED_VECTOR_PADDSUBPD_INVERT);
   auto InvertedSource = _VXor(Size, ElementSize, Src2, ConstantEOR);
   return _VFAdd(Size, ElementSize, Src1, InvertedSource);
 }
@@ -3858,7 +3858,7 @@ OrderedNode* OpDispatchBuilder::PHMINPOSUWOpImpl(OpcodeArgs) {
   // Setup a vector swizzle
   // Initially load a 64-bit mask of immediates
   // Then zero-extend that to 128-bit mask with the immediates in the lower 16-bits of each element
-  auto ConstantSwizzle = _LoadNamedVectorConstant(Size, FEXCore::IR::NamedVectorConstant::NAMED_VECTOR_INCREMENTAL_U16_INDEX);
+  auto ConstantSwizzle = LoadAndCacheNamedVectorConstant(Size, FEXCore::IR::NamedVectorConstant::NAMED_VECTOR_INCREMENTAL_U16_INDEX);
 
   // We now need to zip the vector sources together to become two uint32x4_t vectors
   // Upper: