diff --git a/Makefile b/Makefile
index 1780121e..145d2bab 100644
--- a/Makefile
+++ b/Makefile
@@ -72,7 +72,8 @@ perf: $(TARGET)
 	# Requires gperftools
 	perf record --call-graph dwarf -o perf.data ./$< -d
 	# Analyze with hotspot (https://github.com/KDAB/hotspot) for example
-	# hotspot perf.data
+	hotspot perf.data
+	rm perf.data
 
 valgrind: $(TARGET)
 	# --gen-suppressions=yes
diff --git a/include/simd/sse_mathfun.h b/include/simd/sse_mathfun.h
index 64188187..69b74d33 100644
--- a/include/simd/sse_mathfun.h
+++ b/include/simd/sse_mathfun.h
@@ -45,7 +45,7 @@ This derived source file is released under the zlib license.
 */
 
 
-#include <emmintrin.h>
+#include <x86intrin.h>
 
 
 /** Generate 1.f without accessing memory */
diff --git a/include/simd/vector.hpp b/include/simd/vector.hpp
index ce1ec3ed..a043f6bf 100644
--- a/include/simd/vector.hpp
+++ b/include/simd/vector.hpp
@@ -1,6 +1,6 @@
 #pragma once
 #include <cstring>
-#include <emmintrin.h>
+#include <x86intrin.h>
 
 
 namespace rack {
diff --git a/src/Core/AudioInterface.cpp b/src/Core/AudioInterface.cpp
index 093ed7a3..92c5edb2 100644
--- a/src/Core/AudioInterface.cpp
+++ b/src/Core/AudioInterface.cpp
@@ -132,7 +132,6 @@ struct AudioInterface : Module {
 		if (port.active && port.numInputs > 0) {
 			// Wait until inputs are present
 			// Give up after a timeout in case the audio device is being unresponsive.
-			APP->engine->yieldWorkers();
 			std::unique_lock<std::mutex> lock(port.engineMutex);
 			auto cond = [&] {
 				return (!port.inputBuffer.empty());
@@ -182,11 +181,12 @@ struct AudioInterface : Module {
 			if (outputBuffer.full()) {
 				// Wait until enough outputs are consumed
 				// Give up after a timeout in case the audio device is being unresponsive.
-				APP->engine->yieldWorkers();
-				std::unique_lock<std::mutex> lock(port.engineMutex);
 				auto cond = [&] {
 					return (port.outputBuffer.size() < (size_t) port.blockSize);
 				};
+				if (!cond())
+					APP->engine->yieldWorkers();
+				std::unique_lock<std::mutex> lock(port.engineMutex);
 				auto timeout = std::chrono::milliseconds(200);
 				if (port.engineCv.wait_for(lock, timeout, cond)) {
 					// Push converted output
diff --git a/src/engine/Engine.cpp b/src/engine/Engine.cpp
index 6b14c05f..4b1b6cf1 100644
--- a/src/engine/Engine.cpp
+++ b/src/engine/Engine.cpp
@@ -9,8 +9,7 @@
 #include <condition_variable>
 #include <mutex>
 #include <atomic>
-#include <xmmintrin.h>
-#include <pmmintrin.h>
+#include <x86intrin.h>
 
 
 namespace rack {
@@ -90,7 +89,9 @@ struct SpinBarrier {
 			count = 0;
 		}
 		else {
-			while (count != 0);
+			while (count != 0) {
+				_mm_pause();
+			}
 		}
 	}
 };
@@ -127,6 +128,7 @@ struct HybridBarrier {
 		while (!yield) {
 			if (count == 0)
 				return;
+			_mm_pause();
 		}
 
 		// Wait on mutex