diff --git a/FEXCore/Scripts/json_ir_generator.py b/FEXCore/Scripts/json_ir_generator.py
index 425c966e85..4a53d074a7 100755
--- a/FEXCore/Scripts/json_ir_generator.py
+++ b/FEXCore/Scripts/json_ir_generator.py
@@ -223,7 +223,7 @@ def parse_ops(ops):
                         (OpArg.Type == "GPR" or
                         OpArg.Type == "GPRPair" or
                         OpArg.Type == "FPR" or
-                        OpArg.Type == "PR")):
+                        OpArg.Type == "PRED")):
                         OpDef.EmitValidation.append(f"GetOpRegClass({ArgName}) == InvalidClass || WalkFindRegClass({ArgName}) == {OpArg.Type}Class")
 
                     OpArg.Name = ArgName
diff --git a/FEXCore/Source/Interface/Core/OpcodeDispatcher.cpp b/FEXCore/Source/Interface/Core/OpcodeDispatcher.cpp
index 37efb44008..bbd5498c43 100644
--- a/FEXCore/Source/Interface/Core/OpcodeDispatcher.cpp
+++ b/FEXCore/Source/Interface/Core/OpcodeDispatcher.cpp
@@ -4312,7 +4312,7 @@ Ref OpDispatchBuilder::LoadSource_WithOpSize(RegisterClassType Class, const X86T
       Ref MemSrc = LoadEffectiveAddress(A, true);
       if (CTX->HostFeatures.SupportsSVE128 || CTX->HostFeatures.SupportsSVE256) {
         // Using SVE we can load this with a single instruction.
-        auto PReg = _InitPredicate(OpSize::i16Bit, FEXCore::ToUnderlying(ARMEmitter::PredicatePattern::SVE_VL5));
+        auto PReg = InitPredicateCached(OpSize::i16Bit, ARMEmitter::PredicatePattern::SVE_VL5);
         return _LoadMemPredicate(OpSize::i128Bit, OpSize::i16Bit, PReg, MemSrc);
       } else {
         // For X87 extended doubles, Split the load.
@@ -4446,7 +4446,7 @@ void OpDispatchBuilder::StoreResult_WithOpSize(FEXCore::IR::RegisterClassType Cl
   if (OpSize == OpSize::f80Bit) {
     Ref MemStoreDst = LoadEffectiveAddress(A, true);
     if (CTX->HostFeatures.SupportsSVE128 || CTX->HostFeatures.SupportsSVE256) {
-      auto PReg = _InitPredicate(OpSize::i16Bit, FEXCore::ToUnderlying(ARMEmitter::PredicatePattern::SVE_VL5));
+      auto PReg = InitPredicateCached(OpSize::i16Bit, ARMEmitter::PredicatePattern::SVE_VL5);
       _StoreMemPredicate(OpSize::i128Bit, OpSize::i16Bit, Src, PReg, MemStoreDst);
     } else {
       // For X87 extended doubles, split before storing
diff --git a/FEXCore/Source/Interface/Core/OpcodeDispatcher.h b/FEXCore/Source/Interface/Core/OpcodeDispatcher.h
index ec6589c9c7..61036a2b57 100644
--- a/FEXCore/Source/Interface/Core/OpcodeDispatcher.h
+++ b/FEXCore/Source/Interface/Core/OpcodeDispatcher.h
@@ -125,6 +125,9 @@ class OpDispatchBuilder final : public IREmitter {
 
     // Need to clear any named constants that were cached.
     ClearCachedNamedConstants();
+
+    // Clear predicate cache for x87 ldst
+    ResetInitPredicateCache();
   }
 
   IRPair<IROp_Jump> Jump() {
diff --git a/FEXCore/Source/Interface/IR/IREmitter.cpp b/FEXCore/Source/Interface/IR/IREmitter.cpp
index 0850187b1c..95cb2e73dd 100644
--- a/FEXCore/Source/Interface/IR/IREmitter.cpp
+++ b/FEXCore/Source/Interface/IR/IREmitter.cpp
@@ -41,6 +41,7 @@ FEXCore::IR::RegisterClassType IREmitter::WalkFindRegClass(Ref Node) {
   case FPRClass:
   case GPRFixedClass:
   case FPRFixedClass:
+  case PREDClass:
   case InvalidClass: return Class;
   default: break;
   }
diff --git a/FEXCore/Source/Interface/IR/IREmitter.h b/FEXCore/Source/Interface/IR/IREmitter.h
index 0cfc4027be..c5af4efdd3 100644
--- a/FEXCore/Source/Interface/IR/IREmitter.h
+++ b/FEXCore/Source/Interface/IR/IREmitter.h
@@ -1,6 +1,7 @@
 // SPDX-License-Identifier: MIT
 #pragma once
 
+#include "CodeEmitter/Emitter.h"
 #include "Interface/IR/IR.h"
 #include "Interface/IR/IntrusiveIRList.h"
 
@@ -9,9 +10,9 @@
 
 #include <FEXCore/Utils/LogManager.h>
 #include <FEXCore/fextl/vector.h>
+#include <FEXCore/fextl/unordered_map.h>
 
 #include <algorithm>
-#include <new>
 #include <stdint.h>
 #include <string.h>
 
@@ -45,6 +46,37 @@ class IREmitter {
   }
   void ResetWorkingList();
 
+  // Predicate Cache Implementation
+  // This lives here rather than OpcodeDispatcher because x87StackOptimization Pass
+  // also needs it.
+  struct PredicateKey {
+    ARMEmitter::PredicatePattern Pattern;
+    OpSize Size;
+    bool operator==(const PredicateKey& rhs) const = default;
+  };
+
+  struct PredicateKeyHash {
+    size_t operator()(const PredicateKey& key) const {
+      return FEXCore::ToUnderlying(key.Pattern) + (FEXCore::ToUnderlying(key.Size) * FEXCore::ToUnderlying(OpSize::iInvalid));
+    }
+  };
+  fextl::unordered_map<PredicateKey, Ref, PredicateKeyHash> InitPredicateCache;
+
+  Ref InitPredicateCached(OpSize Size, ARMEmitter::PredicatePattern Pattern) {
+    PredicateKey Key {Pattern, Size};
+    auto ValIt = InitPredicateCache.find(Key);
+    if (ValIt == InitPredicateCache.end()) {
+      auto Predicate = _InitPredicate(Size, static_cast<uint8_t>(FEXCore::ToUnderlying(Pattern)));
+      InitPredicateCache[Key] = Predicate;
+      return Predicate;
+    }
+    return ValIt->second;
+  }
+
+  void ResetInitPredicateCache() {
+    InitPredicateCache.clear();
+  }
+
   /**
    * @name IR allocation routines
    *
diff --git a/FEXCore/Source/Interface/IR/Passes/x87StackOptimizationPass.cpp b/FEXCore/Source/Interface/IR/Passes/x87StackOptimizationPass.cpp
index a34b4fdd3a..9c45c769a0 100644
--- a/FEXCore/Source/Interface/IR/Passes/x87StackOptimizationPass.cpp
+++ b/FEXCore/Source/Interface/IR/Passes/x87StackOptimizationPass.cpp
@@ -824,7 +824,7 @@ void X87StackOptimization::Run(IREmitter* Emit) {
             }
             if (Op->StoreSize == OpSize::f80Bit) { // Part of code from StoreResult_WithOpSize()
               if (Features.SupportsSVE128 || Features.SupportsSVE256) {
-                auto PReg = IREmit->_InitPredicate(OpSize::i16Bit, FEXCore::ToUnderlying(ARMEmitter::PredicatePattern::SVE_VL5));
+                auto PReg = IREmit->InitPredicateCached(OpSize::i16Bit, ARMEmitter::PredicatePattern::SVE_VL5);
                 IREmit->_StoreMemPredicate(OpSize::i128Bit, OpSize::i16Bit, StackNode, PReg, AddrNode);
               } else {
                 // For X87 extended doubles, split before storing