diff --git a/.github/workflows/build.yml b/.github/workflows/build.yml
index 2442548..cad0d0f 100644
--- a/.github/workflows/build.yml
+++ b/.github/workflows/build.yml
@@ -17,7 +17,7 @@ jobs:
       uses: actions/checkout@v3
 
     - name: "Setup node.js"
-      uses: actions/setup-node@v1
+      uses: actions/setup-node@v3
       with:
         node-version: "14"
 
diff --git a/src/asmjit/core.h b/src/asmjit/core.h
index 5a07269..a9e4df7 100644
--- a/src/asmjit/core.h
+++ b/src/asmjit/core.h
@@ -454,7 +454,8 @@ namespace asmjit {
 //!   JitRuntime rt;                    // Runtime specialized for JIT code execution.
 //!
 //!   CodeHolder code;                  // Holds code and relocation information.
-//!   code.init(rt.environment());      // Initialize code to match the JIT environment.
+//!   code.init(rt.environment(),       // Initialize code to match the JIT environment.
+//!             rt.cpuFeatures());
 //!
 //!   x86::Assembler a(&code);          // Create and attach x86::Assembler to code.
 //!   a.mov(x86::eax, 1);               // Move one to eax register.
@@ -599,9 +600,10 @@ namespace asmjit {
 //! int main() {
 //!   // Create a custom environment that matches the current host environment.
 //!   Environment env = Environment::host();
+//!   CpuFeatures cpuFeatures = CpuInfo::host().features();
 //!
 //!   CodeHolder code;                  // Create a CodeHolder.
-//!   code.init(env);                   // Initialize CodeHolder with environment.
+//!   code.init(env, cpuFeatures);      // Initialize CodeHolder with environment.
 //!
 //!   x86::Assembler a(&code);          // Create and attach x86::Assembler to `code`.
 //!
@@ -722,10 +724,11 @@ namespace asmjit {
 //!
 //! void initializeCodeHolder(CodeHolder& code) {
 //!   Environment env = Environment::host();
+//!   CpuFeatures cpuFeatures = CpuInfo::host().features();
 //!   uint64_t baseAddress = uint64_t(0x1234);
 //!
 //!   // initialize CodeHolder with environment and custom base address.
-//!   code.init(env, baseAddress);
+//!   code.init(env, cpuFeatures, baseAddress);
 //! }
 //! ```
 //!
@@ -1346,7 +1349,8 @@ namespace asmjit {
 //!   FileLogger logger(stdout);   // Logger should always survive CodeHolder.
 //!
 //!   CodeHolder code;             // Holds code and relocation information.
-//!   code.init(rt.environment()); // Initialize to the same arch as JIT runtime.
+//!   code.init(rt.environment(),  // Initialize code to match the JIT environment.
+//!             rt.cpuFeatures());
 //!   code.setLogger(&logger);     // Attach the `logger` to `code` holder.
 //!
 //!   // ... code as usual, everything emitted will be logged to `stdout` ...
@@ -1369,7 +1373,8 @@ namespace asmjit {
 //!   StringLogger logger;         // Logger should always survive CodeHolder.
 //!
 //!   CodeHolder code;             // Holds code and relocation information.
-//!   code.init(rt.environment()); // Initialize to the same arch as JIT runtime.
+//!   code.init(rt.environment(),  // Initialize code to match the JIT environment.
+//!             rt.cpuFeatures());
 //!   code.setLogger(&logger);     // Attach the `logger` to `code` holder.
 //!
 //!   // ... code as usual, logging will be concatenated to logger string  ...
@@ -1560,7 +1565,7 @@ namespace asmjit {
 //!   MyErrorHandler myErrorHandler;
 //!   CodeHolder code;
 //!
-//!   code.init(rt.environment());
+//!   code.init(rt.environment(), rt.cpuFeatures());
 //!   code.setErrorHandler(&myErrorHandler);
 //!
 //!   x86::Assembler a(&code);
diff --git a/src/asmjit/core/codeholder.cpp b/src/asmjit/core/codeholder.cpp
index cf763cf..3f7f5f5 100644
--- a/src/asmjit/core/codeholder.cpp
+++ b/src/asmjit/core/codeholder.cpp
@@ -74,6 +74,7 @@ static void CodeHolder_resetInternal(CodeHolder* self, ResetPolicy resetPolicy)
 
   // Reset everything into its construction state.
   self->_environment.reset();
+  self->_cpuFeatures.reset();
   self->_baseAddress = Globals::kNoBaseAddress;
   self->_logger = nullptr;
   self->_errorHandler = nullptr;
@@ -118,6 +119,7 @@ static void CodeHolder_onSettingsUpdated(CodeHolder* self) noexcept {
 
 CodeHolder::CodeHolder(const Support::Temporary* temporary) noexcept
   : _environment(),
+    _cpuFeatures{},
     _baseAddress(Globals::kNoBaseAddress),
     _logger(nullptr),
     _errorHandler(nullptr),
@@ -143,6 +145,10 @@ inline void CodeHolder_setSectionDefaultName(
 }
 
 Error CodeHolder::init(const Environment& environment, uint64_t baseAddress) noexcept {
+  return init(environment, CpuFeatures{}, baseAddress);
+}
+
+Error CodeHolder::init(const Environment& environment, const CpuFeatures& cpuFeatures, uint64_t baseAddress) noexcept {
   // Cannot reinitialize if it's locked or there is one or more emitter attached.
   if (isInitialized())
     return DebugUtils::errored(kErrorAlreadyInitialized);
@@ -172,6 +178,7 @@ Error CodeHolder::init(const Environment& environment, uint64_t baseAddress) noe
   }
   else {
     _environment = environment;
+    _cpuFeatures = cpuFeatures;
     _baseAddress = baseAddress;
     return kErrorOk;
   }
diff --git a/src/asmjit/core/codeholder.h b/src/asmjit/core/codeholder.h
index 6ed2ddf..d663d3e 100644
--- a/src/asmjit/core/codeholder.h
+++ b/src/asmjit/core/codeholder.h
@@ -640,6 +640,8 @@ public:
 
   //! Environment information.
   Environment _environment;
+  //! CPU features of the target architecture.
+  CpuFeatures _cpuFeatures;
   //! Base address or \ref Globals::kNoBaseAddress.
   uint64_t _baseAddress;
 
@@ -698,6 +700,8 @@ public:
 
   //! Initializes CodeHolder to hold code described by the given `environment` and `baseAddress`.
   ASMJIT_API Error init(const Environment& environment, uint64_t baseAddress = Globals::kNoBaseAddress) noexcept;
+  //! Initializes CodeHolder to hold code described by the given `environment`, `cpuFeatures`, and `baseAddress`.
+  ASMJIT_API Error init(const Environment& environment, const CpuFeatures& cpuFeatures, uint64_t baseAddress = Globals::kNoBaseAddress) noexcept;
   //! Detaches all code-generators attached and resets the `CodeHolder`.
   ASMJIT_API void reset(ResetPolicy resetPolicy = ResetPolicy::kSoft) noexcept;
 
@@ -736,6 +740,9 @@ public:
   //! Returns the target sub-architecture.
   inline SubArch subArch() const noexcept { return environment().subArch(); }
 
+  //! Returns the minimum CPU features of the target architecture.
+  inline const CpuFeatures& cpuFeatures() const noexcept { return _cpuFeatures; }
+
   //! Tests whether a static base-address is set.
   inline bool hasBaseAddress() const noexcept { return _baseAddress != Globals::kNoBaseAddress; }
   //! Returns a static base-address or \ref Globals::kNoBaseAddress, if not set.
diff --git a/src/asmjit/core/errorhandler.h b/src/asmjit/core/errorhandler.h
index 5151d43..3b0096f 100644
--- a/src/asmjit/core/errorhandler.h
+++ b/src/asmjit/core/errorhandler.h
@@ -61,7 +61,7 @@ class BaseEmitter;
 //!   SimpleErrorHandler eh;
 //!
 //!   CodeHolder code;
-//!   code.init(rt.environment());
+//!   code.init(rt.environment(), rt.cpuFeatures());
 //!   code.setErrorHandler(&eh);
 //!
 //!   // Try to emit instruction that doesn't exist.
@@ -117,7 +117,7 @@ class BaseEmitter;
 //!   ThrowableErrorHandler eh;
 //!
 //!   CodeHolder code;
-//!   code.init(rt.environment());
+//!   code.init(rt.environment(), rt.cpuFeatures());
 //!   code.setErrorHandler(&eh);
 //!
 //!   x86::Assembler a(&code);
@@ -166,7 +166,7 @@ class BaseEmitter;
 //!   LongJmpErrorHandler eh;
 //!
 //!   CodeHolder code;
-//!   code.init(rt.rt.environment());
+//!   code.init(rt.environment(), rt.cpuFeatures());
 //!   code.setErrorHandler(&eh);
 //!
 //!   x86::Assembler a(&code);
diff --git a/src/asmjit/core/jitruntime.cpp b/src/asmjit/core/jitruntime.cpp
index 491c204..be1c2f4 100644
--- a/src/asmjit/core/jitruntime.cpp
+++ b/src/asmjit/core/jitruntime.cpp
@@ -15,6 +15,7 @@ JitRuntime::JitRuntime(const JitAllocator::CreateParams* params) noexcept
   : _allocator(params) {
   _environment = Environment::host();
   _environment.setObjectFormat(ObjectFormat::kJIT);
+  _cpuFeatures = CpuInfo::host().features();
 }
 
 JitRuntime::~JitRuntime() noexcept {}
diff --git a/src/asmjit/core/target.cpp b/src/asmjit/core/target.cpp
index fef025d..cbc6ab5 100644
--- a/src/asmjit/core/target.cpp
+++ b/src/asmjit/core/target.cpp
@@ -8,7 +8,9 @@
 
 ASMJIT_BEGIN_NAMESPACE
 
-Target::Target() noexcept : _environment() {}
+Target::Target() noexcept
+  : _environment{},
+    _cpuFeatures{} {}
 Target::~Target() noexcept {}
 
 ASMJIT_END_NAMESPACE
diff --git a/src/asmjit/core/target.h b/src/asmjit/core/target.h
index 23b0c62..322a338 100644
--- a/src/asmjit/core/target.h
+++ b/src/asmjit/core/target.h
@@ -7,6 +7,7 @@
 #define ASMJIT_CORE_TARGET_H_INCLUDED
 
 #include "../core/archtraits.h"
+#include "../core/cpuinfo.h"
 #include "../core/func.h"
 
 ASMJIT_BEGIN_NAMESPACE
@@ -22,6 +23,8 @@ public:
 
   //! Target environment information.
   Environment _environment;
+  //! Target CPU features.
+  CpuFeatures _cpuFeatures;
 
   //! \name Construction & Destruction
   //! \{
@@ -43,6 +46,9 @@ public:
   //! Returns the target sub-architecture.
   inline SubArch subArch() const noexcept { return _environment.subArch(); }
 
+  //! Returns target CPU features.
+  inline const CpuFeatures& cpuFeatures() const noexcept { return _cpuFeatures; }
+
   //! \}
 };
 
diff --git a/src/asmjit/x86/x86assembler.h b/src/asmjit/x86/x86assembler.h
index dbffae6..dde27f4 100644
--- a/src/asmjit/x86/x86assembler.h
+++ b/src/asmjit/x86/x86assembler.h
@@ -38,7 +38,8 @@ ASMJIT_BEGIN_SUB_NAMESPACE(x86)
 //!   JitRuntime rt;                    // Create a runtime specialized for JIT.
 //!   CodeHolder code;                  // Create a CodeHolder.
 //!
-//!   code.init(rt.environment());      // Initialize code to match the JIT environment.
+//!   code.init(rt.environment(),       // Initialize code to match the JIT environment.
+//!             rt.cpuFeatures());
 //!   x86::Assembler a(&code);          // Create and attach x86::Assembler to code.
 //!
 //!   // Decide between 32-bit CDECL, WIN64, and SysV64 calling conventions:
@@ -131,7 +132,8 @@ ASMJIT_BEGIN_SUB_NAMESPACE(x86)
 //!   JitRuntime rt;                    // Create a runtime specialized for JIT.
 //!   CodeHolder code;                  // Create a CodeHolder.
 //!
-//!   code.init(rt.environment());      // Initialize code to match the JIT environment.
+//!   code.init(rt.environment(),       // Initialize code to match the JIT environment.
+//!             rt.cpuFeatures());
 //!   x86::Assembler a(&code);          // Create and attach x86::Assembler to code.
 //!
 //!   // Enable strict validation.
@@ -187,7 +189,8 @@ ASMJIT_BEGIN_SUB_NAMESPACE(x86)
 //!   JitRuntime rt;                    // Create a runtime specialized for JIT.
 //!   CodeHolder code;                  // Create a CodeHolder.
 //!
-//!   code.init(rt.environment());      // Initialize code to match the JIT environment.
+//!   code.init(rt.environment(),       // Initialize code to match the JIT environment.
+//!             rt.cpuFeatures());
 //!   x86::Assembler a(&code);          // Create and attach x86::Assembler to code.
 //!
 //!   // Let's get these registers from x86::Assembler.
@@ -346,7 +349,8 @@ ASMJIT_BEGIN_SUB_NAMESPACE(x86)
 //!   JitRuntime rt;                    // Create JIT Runtime.
 //!   CodeHolder code;                  // Create a CodeHolder.
 //!
-//!   code.init(rt.environment());      // Initialize code to match the JIT environment.
+//!   code.init(rt.environment(),       // Initialize code to match the JIT environment.
+//!             rt.cpuFeatures());
 //!   x86::Assembler a(&code);          // Create and attach x86::Assembler to code.
 //!
 //!   // Decide which registers will be mapped to function arguments. Try changing
@@ -453,7 +457,8 @@ ASMJIT_BEGIN_SUB_NAMESPACE(x86)
 //!   JitRuntime rt;                    // Create a runtime specialized for JIT.
 //!   CodeHolder code;                  // Create a CodeHolder.
 //!
-//!   code.init(rt.environment());      // Initialize code to match the JIT environment.
+//!   code.init(rt.environment(),       // Initialize code to match the JIT environment.
+//!             rt.cpuFeatures());
 //!   x86::Assembler a(&code);          // Create and attach x86::Assembler to code.
 //!
 //!   // Let's get these registers from x86::Assembler.
diff --git a/src/asmjit/x86/x86builder.h b/src/asmjit/x86/x86builder.h
index 5671c9d..4c56cba 100644
--- a/src/asmjit/x86/x86builder.h
+++ b/src/asmjit/x86/x86builder.h
@@ -46,7 +46,8 @@ ASMJIT_BEGIN_SUB_NAMESPACE(x86)
 //!   JitRuntime rt;                    // Create JIT Runtime.
 //!   CodeHolder code;                  // Create a CodeHolder.
 //!
-//!   code.init(rt.environment());      // Initialize code to match the JIT environment.
+//!   code.init(rt.environment(),       // Initialize code to match the JIT environment.
+//!             rt.cpuFeatures());
 //!   x86::Builder cb(&code);           // Create and attach x86::Builder to `code`.
 //!
 //!   // Decide which registers will be mapped to function arguments. Try changing registers
diff --git a/src/asmjit/x86/x86compiler.h b/src/asmjit/x86/x86compiler.h
index d89aea0..8bb9a43 100644
--- a/src/asmjit/x86/x86compiler.h
+++ b/src/asmjit/x86/x86compiler.h
@@ -38,7 +38,8 @@ ASMJIT_BEGIN_SUB_NAMESPACE(x86)
 //!   JitRuntime rt;                    // Runtime specialized for JIT code execution.
 //!   CodeHolder code;                  // Holds code and relocation information.
 //!
-//!   code.init(rt.environment());      // Initialize code to match the JIT environment.
+//!   code.init(rt.environment(),       // Initialize code to match the JIT environment.
+//!             rt.cpuFeatures());
 //!   x86::Compiler cc(&code);          // Create and attach x86::Compiler to code.
 //!
 //!   cc.addFunc(FuncSignatureT<int>());// Begin a function of `int fn(void)` signature.
@@ -82,7 +83,8 @@ ASMJIT_BEGIN_SUB_NAMESPACE(x86)
 //!   JitRuntime rt;                    // Runtime specialized for JIT code execution.
 //!   CodeHolder code;                  // Holds code and relocation information.
 //!
-//!   code.init(rt.environment());      // Initialize code to match the JIT environment.
+//!   code.init(rt.environment(),       // Initialize code to match the JIT environment.
+//!             rt.cpuFeatures());
 //!   x86::Compiler cc(&code);          // Create and attach x86::Compiler to code.
 //!
 //!   FuncNode* funcNode = cc.addFunc(  // Begin the function of the following signature:
@@ -164,7 +166,8 @@ ASMJIT_BEGIN_SUB_NAMESPACE(x86)
 //!   JitRuntime rt;                    // Runtime specialized for JIT code execution.
 //!   CodeHolder code;                  // Holds code and relocation information.
 //!
-//!   code.init(rt.environment());      // Initialize code to match the JIT environment.
+//!   code.init(rt.environment(),       // Initialize code to match the JIT environment.
+//!             rt.cpuFeatures());
 //!   x86::Compiler cc(&code);          // Create and attach x86::Compiler to code.
 //!
 //!   FuncNode* funcNode = cc.addFunc(FuncSignatureT<void, void*>());
@@ -222,7 +225,8 @@ ASMJIT_BEGIN_SUB_NAMESPACE(x86)
 //!   JitRuntime rt;                    // Runtime specialized for JIT code execution.
 //!   CodeHolder code;                  // Holds code and relocation information.
 //!
-//!   code.init(rt.environment());      // Initialize code to match the JIT environment.
+//!   code.init(rt.environment(),       // Initialize code to match the JIT environment.
+//!             rt.cpuFeatures());
 //!   x86::Compiler cc(&code);          // Create and attach x86::Compiler to code.
 //!
 //!   FuncNode* funcNode = cc.addFunc(  // Begin of the Fibonacci function, addFunc()
@@ -290,7 +294,8 @@ ASMJIT_BEGIN_SUB_NAMESPACE(x86)
 //!   JitRuntime rt;                    // Runtime specialized for JIT code execution.
 //!   CodeHolder code;                  // Holds code and relocation information.
 //!
-//!   code.init(rt.environment());      // Initialize code to match the JIT environment.
+//!   code.init(rt.environment(),       // Initialize code to match the JIT environment.
+//!             rt.cpuFeatures());
 //!   x86::Compiler cc(&code);          // Create and attach x86::Compiler to code.
 //!
 //!   cc.addFunc(FuncSignatureT<int>());// Create a function that returns int.
diff --git a/src/asmjit/x86/x86rapass.cpp b/src/asmjit/x86/x86rapass.cpp
index 02870d9..4d8f3aa 100644
--- a/src/asmjit/x86/x86rapass.cpp
+++ b/src/asmjit/x86/x86rapass.cpp
@@ -220,9 +220,11 @@ Error RACFGBuilder::onInst(InstNode* inst, InstControlFlow& cf, RAInstBuilder& i
               }
             }
 
-            // Do not use RegMem flag if changing Reg to Mem requires additional CPU feature that may not be enabled.
+            // Do not use RegMem flag if changing Reg to Mem requires a CPU feature that is not available.
             if (rwInfo.rmFeature() && Support::test(flags, RATiedFlags::kUseRM | RATiedFlags::kOutRM)) {
-              flags &= ~(RATiedFlags::kUseRM | RATiedFlags::kOutRM);
+              if (!cc()->code()->cpuFeatures().has(rwInfo.rmFeature())) {
+                flags &= ~(RATiedFlags::kUseRM | RATiedFlags::kOutRM);
+              }
             }
 
             RegGroup group = workReg->group();
diff --git a/test/asmjit_test_compiler.cpp b/test/asmjit_test_compiler.cpp
index f2bc306..293b84e 100644
--- a/test/asmjit_test_compiler.cpp
+++ b/test/asmjit_test_compiler.cpp
@@ -91,7 +91,7 @@ int TestApp::run() {
       CodeHolder code;
       SimpleErrorHandler errorHandler;
 
-      code.init(runtime.environment());
+      code.init(runtime.environment(), runtime.cpuFeatures());
       code.setErrorHandler(&errorHandler);
 
       if (pass != 0) {
diff --git a/test/asmjit_test_emitters.cpp b/test/asmjit_test_emitters.cpp
index 63c14a6..74db996 100644
--- a/test/asmjit_test_emitters.cpp
+++ b/test/asmjit_test_emitters.cpp
@@ -90,7 +90,7 @@ static uint32_t testFunc(JitRuntime& rt, EmitterType emitterType) noexcept {
 #endif
 
   CodeHolder code;
-  code.init(rt.environment());
+  code.init(rt.environment(), rt.cpuFeatures());
 
 #ifndef ASMJIT_NO_LOGGING
   code.setLogger(&logger);
diff --git a/test/asmjit_test_perf.h b/test/asmjit_test_perf.h
index 2ab0038..fb65db5 100644
--- a/test/asmjit_test_perf.h
+++ b/test/asmjit_test_perf.h
@@ -10,7 +10,9 @@
 #include "asmjitutils.h"
 #include "performancetimer.h"
 
-class MyErrorHandler : public asmjit::ErrorHandler {
+namespace asmjit_perf_utils {
+
+class TestErrorHandler : public asmjit::ErrorHandler {
   void handleError(asmjit::Error err, const char* message, asmjit::BaseEmitter* origin) {
     (void)err;
     (void)origin;
@@ -19,10 +21,49 @@ class MyErrorHandler : public asmjit::ErrorHandler {
   }
 };
 
+#ifndef ASMJIT_NO_BUILDER
+template<typename BuilderT, typename FuncT>
+static uint32_t calculateInstructionCount(asmjit::CodeHolder& code, asmjit::Arch arch, const FuncT& func) noexcept {
+  BuilderT builder;
+  TestErrorHandler eh;
+
+  asmjit::Environment env(arch);
+  code.init(env);
+  code.setErrorHandler(&eh);
+  code.attach(&builder);
+  func(builder);
+
+  uint32_t count = 0;
+  asmjit::BaseNode* node = builder.firstNode();
+
+  while (node) {
+    count += uint32_t(node->isInst());
+    node = node->next();
+  }
+
+  return count;
+}
+#endif
+
+static inline double mbps(double duration, uint64_t outputSize) noexcept {
+  if (duration == 0)
+    return 0.0;
+
+  double bytesTotal = double(outputSize);
+  return (bytesTotal * 1000) / (duration * 1024 * 1024);
+}
+
+static inline double mips(double duration, uint64_t instCount) noexcept {
+  if (duration == 0)
+    return 0.0;
+
+  return double(instCount) * 1000.0 / (duration * 1e6);
+}
+
 template<typename EmitterT, typename FuncT>
-static void bench(asmjit::CodeHolder& code, asmjit::Arch arch, uint32_t numIterations, const char* testName, const FuncT& func) noexcept {
+static void bench(asmjit::CodeHolder& code, asmjit::Arch arch, uint32_t numIterations, const char* testName, uint32_t instCount, const FuncT& func) noexcept {
   EmitterT emitter;
-  MyErrorHandler eh;
+  TestErrorHandler eh;
 
   const char* archName = asmjitArchAsString(arch);
   const char* emitterName =
@@ -53,9 +94,20 @@ static void bench(asmjit::CodeHolder& code, asmjit::Arch arch, uint32_t numItera
   }
 
   printf("  [%s] %-9s %-16s | CodeSize:%5llu [B] | Time:%8.4f [ms]", archName, emitterName, testName, (unsigned long long)codeSize, duration);
-  if (codeSize)
+  if (codeSize) {
     printf(" | Speed:%8.3f [MB/s]", mbps(duration, codeSize));
+  }
+  else {
+    printf(" | Speed: N/A           ");
+  }
+
+  if (instCount) {
+    printf(", %8.3f [MI/s]", mips(duration, instCount));
+  }
+
   printf("\n");
 }
 
+} // {asmjit_perf_utils}
+
 #endif // ASMJIT_TEST_PERF_H_INCLUDED
diff --git a/test/asmjit_test_perf_a64.cpp b/test/asmjit_test_perf_a64.cpp
index 4f4aebb..054959b 100644
--- a/test/asmjit_test_perf_a64.cpp
+++ b/test/asmjit_test_perf_a64.cpp
@@ -645,42 +645,50 @@ static void benchmarkA64Function(Arch arch, uint32_t numIterations, const char*
   CodeHolder code;
   printf("%s:\n", description);
 
-  bench<a64::Assembler>(code, arch, numIterations, "[raw]", [&](a64::Assembler& cc) {
+  uint32_t instCount = 0;
+
+#ifndef ASMJIT_NO_BUILDER
+  instCount = asmjit_perf_utils::calculateInstructionCount<a64::Builder>(code, arch, [&](a64::Builder& cc) {
+    emitterFn(cc, false);
+  });
+#endif
+
+  asmjit_perf_utils::bench<a64::Assembler>(code, arch, numIterations, "[raw]", instCount, [&](a64::Assembler& cc) {
     emitterFn(cc, false);
   });
 
-  bench<a64::Assembler>(code, arch, numIterations, "[validated]", [&](a64::Assembler& cc) {
+  asmjit_perf_utils::bench<a64::Assembler>(code, arch, numIterations, "[validated]", instCount, [&](a64::Assembler& cc) {
     cc.addDiagnosticOptions(DiagnosticOptions::kValidateAssembler);
     emitterFn(cc, false);
   });
 
-  bench<a64::Assembler>(code, arch, numIterations, "[prolog/epilog]", [&](a64::Assembler& cc) {
+  asmjit_perf_utils::bench<a64::Assembler>(code, arch, numIterations, "[prolog/epilog]", instCount, [&](a64::Assembler& cc) {
     cc.addDiagnosticOptions(DiagnosticOptions::kValidateAssembler);
     emitterFn(cc, true);
   });
 
 #ifndef ASMJIT_NO_BUILDER
-  bench<a64::Builder>(code, arch, numIterations, "[no-asm]", [&](a64::Builder& cc) {
+  asmjit_perf_utils::bench<a64::Builder>(code, arch, numIterations, "[no-asm]", instCount, [&](a64::Builder& cc) {
     emitterFn(cc, false);
   });
 
-  bench<a64::Builder>(code, arch, numIterations, "[finalized]", [&](a64::Builder& cc) {
+  asmjit_perf_utils::bench<a64::Builder>(code, arch, numIterations, "[finalized]", instCount, [&](a64::Builder& cc) {
     emitterFn(cc, false);
     cc.finalize();
   });
 
-  bench<a64::Builder>(code, arch, numIterations, "[prolog/epilog]", [&](a64::Builder& cc) {
+  asmjit_perf_utils::bench<a64::Builder>(code, arch, numIterations, "[prolog/epilog]", instCount, [&](a64::Builder& cc) {
     emitterFn(cc, true);
     cc.finalize();
   });
 #endif
 
 #ifndef ASMJIT_NO_COMPILER
-  bench<a64::Compiler>(code, arch, numIterations, "[no-asm]", [&](a64::Compiler& cc) {
+  asmjit_perf_utils::bench<a64::Compiler>(code, arch, numIterations, "[no-asm]", instCount, [&](a64::Compiler& cc) {
     emitterFn(cc, true);
   });
 
-  bench<a64::Compiler>(code, arch, numIterations, "[finalized]", [&](a64::Compiler& cc) {
+  asmjit_perf_utils::bench<a64::Compiler>(code, arch, numIterations, "[finalized]", instCount, [&](a64::Compiler& cc) {
     emitterFn(cc, true);
     cc.finalize();
   });
diff --git a/test/asmjit_test_perf_x86.cpp b/test/asmjit_test_perf_x86.cpp
index e2720b7..515de80 100644
--- a/test/asmjit_test_perf_x86.cpp
+++ b/test/asmjit_test_perf_x86.cpp
@@ -4912,42 +4912,49 @@ static void benchmarkX86Function(Arch arch, uint32_t numIterations, const char*
   CodeHolder code;
   printf("%s:\n", description);
 
-  bench<x86::Assembler>(code, arch, numIterations, "[raw]", [&](x86::Assembler& cc) {
+  uint32_t instCount = 0;
+
+#ifndef ASMJIT_NO_BUILDER
+  instCount = asmjit_perf_utils::calculateInstructionCount<x86::Builder>(code, arch, [&](x86::Builder& cc) {
+    emitterFn(cc, false);
+  });
+#endif
+
+  asmjit_perf_utils::bench<x86::Assembler>(code, arch, numIterations, "[raw]", instCount, [&](x86::Assembler& cc) {
     emitterFn(cc, false);
   });
 
-  bench<x86::Assembler>(code, arch, numIterations, "[validated]", [&](x86::Assembler& cc) {
+  asmjit_perf_utils::bench<x86::Assembler>(code, arch, numIterations, "[validated]", instCount, [&](x86::Assembler& cc) {
     cc.addDiagnosticOptions(DiagnosticOptions::kValidateAssembler);
     emitterFn(cc, false);
   });
 
-  bench<x86::Assembler>(code, arch, numIterations, "[prolog/epilog]", [&](x86::Assembler& cc) {
-    cc.addDiagnosticOptions(DiagnosticOptions::kValidateAssembler);
+  asmjit_perf_utils::bench<x86::Assembler>(code, arch, numIterations, "[prolog/epilog]", instCount, [&](x86::Assembler& cc) {
     emitterFn(cc, true);
   });
 
 #ifndef ASMJIT_NO_BUILDER
-  bench<x86::Builder>(code, arch, numIterations, "[no-asm]", [&](x86::Builder& cc) {
+  asmjit_perf_utils::bench<x86::Builder>(code, arch, numIterations, "[no-asm]", instCount, [&](x86::Builder& cc) {
     emitterFn(cc, false);
   });
 
-  bench<x86::Builder>(code, arch, numIterations, "[finalized]", [&](x86::Builder& cc) {
+  asmjit_perf_utils::bench<x86::Builder>(code, arch, numIterations, "[finalized]", instCount, [&](x86::Builder& cc) {
     emitterFn(cc, false);
     cc.finalize();
   });
 
-  bench<x86::Builder>(code, arch, numIterations, "[prolog/epilog]", [&](x86::Builder& cc) {
+  asmjit_perf_utils::bench<x86::Builder>(code, arch, numIterations, "[prolog/epilog]", instCount, [&](x86::Builder& cc) {
     emitterFn(cc, true);
     cc.finalize();
   });
 #endif
 
 #ifndef ASMJIT_NO_COMPILER
-  bench<x86::Compiler>(code, arch, numIterations, "[no-asm]", [&](x86::Compiler& cc) {
+  asmjit_perf_utils::bench<x86::Compiler>(code, arch, numIterations, "[no-asm]", instCount, [&](x86::Compiler& cc) {
     emitterFn(cc, true);
   });
 
-  bench<x86::Compiler>(code, arch, numIterations, "[finalized]", [&](x86::Compiler& cc) {
+  asmjit_perf_utils::bench<x86::Compiler>(code, arch, numIterations, "[finalized]", instCount, [&](x86::Compiler& cc) {
     emitterFn(cc, true);
     cc.finalize();
   });
diff --git a/test/performancetimer.h b/test/performancetimer.h
index baa0479..c7a8ebe 100644
--- a/test/performancetimer.h
+++ b/test/performancetimer.h
@@ -30,12 +30,4 @@ public:
   }
 };
 
-static inline double mbps(double duration, uint64_t outputSize) noexcept {
-  if (duration == 0)
-    return 0.0;
-
-  double bytesTotal = double(outputSize);
-  return (bytesTotal * 1000) / (duration * 1024 * 1024);
-}
-
 #endif // PERFORMANCETIMER_H_INCLUDED