diff --git a/src/common/heap_array.h b/src/common/heap_array.h
index 14a7b8254..163f0183d 100644
--- a/src/common/heap_array.h
+++ b/src/common/heap_array.h
@@ -279,6 +279,8 @@ public:
     m_size = 0;
   }
 
+  void assign(const std::span<const T> data) { assign(data.data(), data.size()); }
+
   void assign(const T* begin, const T* end)
   {
     const size_t size = reinterpret_cast<const char*>(end) - reinterpret_cast<const char*>(begin);
diff --git a/src/util/gpu_device.cpp b/src/util/gpu_device.cpp
index 8285d8ce4..13ee8644a 100644
--- a/src/util/gpu_device.cpp
+++ b/src/util/gpu_device.cpp
@@ -1780,6 +1780,13 @@ std::unique_ptr<GPUShader> GPUDevice::TranspileAndCreateShaderFromSource(
   GPUShaderStage stage, GPUShaderLanguage source_language, std::string_view source, const char* entry_point,
   GPUShaderLanguage target_language, u32 target_version, DynamicHeapArray<u8>* out_binary, Error* error)
 {
+  // Currently, entry points must be "main". TODO: rename the entry point in the SPIR-V.
+  if (std::strcmp(entry_point, "main") != 0)
+  {
+    Error::SetStringView(error, "Entry point must be main.");
+    return {};
+  }
+
   // Disable optimization when targeting OpenGL GLSL, otherwise, the name-based linking will fail.
   const bool optimization =
     (!m_debug_device && target_language != GPUShaderLanguage::GLSL && target_language != GPUShaderLanguage::GLSLES);
@@ -1827,7 +1834,14 @@ std::unique_ptr<GPUShader> GPUDevice::TranspileAndCreateShaderFromSource(
   if (!TranslateVulkanSpvToLanguage(spv, stage, target_language, target_version, &dest_source, error))
     return {};
 
-  // TODO: MSL needs entry point suffixed.
+#ifdef __APPLE__
+  // MSL converter suffixes 0.
+  if (target_language == GPUShaderLanguage::MSL)
+  {
+    return CreateShaderFromSource(stage, target_language, dest_source,
+                                  TinyString::from_format("{}0", entry_point).c_str(), out_binary, error);
+  }
+#endif
 
   return CreateShaderFromSource(stage, target_language, dest_source, entry_point, out_binary, error);
 }
diff --git a/src/util/metal_device.h b/src/util/metal_device.h
index 9c5c0415c..0257eea9a 100644
--- a/src/util/metal_device.h
+++ b/src/util/metal_device.h
@@ -317,14 +317,12 @@ private:
   void SetFeatures(FeatureMask disabled_features);
   bool LoadShaders();
 
+  std::unique_ptr<GPUShader> CreateShaderFromMSL(GPUShaderStage stage, std::string_view source,
+                                                 std::string_view entry_point, Error* error);
   id<MTLFunction> GetFunctionFromLibrary(id<MTLLibrary> library, NSString* name);
   id<MTLComputePipelineState> CreateComputePipeline(id<MTLFunction> function, NSString* name);
   ClearPipelineConfig GetCurrentClearPipelineConfig() const;
   id<MTLRenderPipelineState> GetClearDepthPipeline(const ClearPipelineConfig& config);
-
-  std::unique_ptr<GPUShader> CreateShaderFromMSL(GPUShaderStage stage, std::string_view source,
-                                                 std::string_view entry_point, Error* error);
-
   id<MTLDepthStencilState> GetDepthState(const GPUPipeline::DepthState& ds);
 
   void CreateCommandBuffer();
diff --git a/src/util/metal_device.mm b/src/util/metal_device.mm
index d8a3626cd..99bae4ab2 100644
--- a/src/util/metal_device.mm
+++ b/src/util/metal_device.mm
@@ -24,6 +24,18 @@ Log_SetChannel(MetalDevice);
 
 // TODO: Disable hazard tracking and issue barriers explicitly.
 
+// Used for shader "binaries".
+namespace {
+struct MetalShaderBinaryHeader
+{
+  u32 entry_point_offset;
+  u32 entry_point_length;
+  u32 source_offset;
+  u32 source_length;
+};
+static_assert(sizeof(MetalShaderBinaryHeader) == 16);
+} // namespace
+
 // Looking across a range of GPUs, the optimal copy alignment for Vulkan drivers seems
 // to be between 1 (AMD/NV) and 64 (Intel). So, we'll go with 64 here.
 static constexpr u32 TEXTURE_UPLOAD_ALIGNMENT = 64;
@@ -648,39 +660,55 @@ std::unique_ptr<GPUShader> MetalDevice::CreateShaderFromMSL(GPUShaderStage stage
 std::unique_ptr<GPUShader> MetalDevice::CreateShaderFromBinary(GPUShaderStage stage, std::span<const u8> data,
                                                                Error* error)
 {
-  const std::string_view str_data(reinterpret_cast<const char*>(data.data()), data.size());
-  return CreateShaderFromMSL(stage, str_data, "main0", error);
+  if (data.size() < sizeof(MetalShaderBinaryHeader))
+  {
+    Error::SetStringView(error, "Invalid header.");
+    return {};
+  }
+
+  // Need to copy for alignment reasons.
+  MetalShaderBinaryHeader hdr;
+  std::memcpy(&hdr, data.data(), sizeof(hdr));
+  if (static_cast<size_t>(hdr.entry_point_offset) + static_cast<size_t>(hdr.entry_point_length) > data.size() ||
+      static_cast<size_t>(hdr.source_offset) + static_cast<size_t>(hdr.source_length) > data.size())
+  {
+    Error::SetStringView(error, "Out of range fields in header.");
+    return {};
+  }
+
+  const std::string_view entry_point(reinterpret_cast<const char*>(data.data() + hdr.entry_point_offset),
+                                     hdr.entry_point_length);
+  const std::string source(reinterpret_cast<const char*>(data.data() + hdr.source_offset), hdr.source_length);
+  return CreateShaderFromMSL(stage, source, entry_point, error);
 }
 
 std::unique_ptr<GPUShader> MetalDevice::CreateShaderFromSource(GPUShaderStage stage, GPUShaderLanguage language,
                                                                std::string_view source, const char* entry_point,
                                                                DynamicHeapArray<u8>* out_binary, Error* error)
 {
-  static constexpr bool dump_shaders = false;
-
-  DynamicHeapArray<u8> spv;
-  if (!CompileGLSLShaderToVulkanSpv(stage, language, source, entry_point, !m_debug_device, false, &spv, error))
-    return {};
-
-  std::string msl;
-  if (!TranslateVulkanSpvToLanguage(spv.cspan(), stage, GPUShaderLanguage::MSL, 230, &msl, error))
-    return {};
-
-  if constexpr (dump_shaders)
+  if (language != GPUShaderLanguage::MSL)
   {
-    static unsigned s_next_id = 0;
-    ++s_next_id;
-    DumpShader(s_next_id, "_input", source);
-    DumpShader(s_next_id, "_msl", msl);
+    return TranspileAndCreateShaderFromSource(stage, language, source, entry_point, GPUShaderLanguage::MSL,
+                                              m_render_api_version, out_binary, error);
   }
 
+  // Source is the "binary" here, since Metal doesn't allow us to access the bytecode :(
+  const std::span<const u8> msl(reinterpret_cast<const u8*>(source.data()), source.size());
   if (out_binary)
   {
-    out_binary->resize(msl.size());
-    std::memcpy(out_binary->data(), msl.data(), msl.size());
+    MetalShaderBinaryHeader hdr;
+    hdr.entry_point_offset = sizeof(MetalShaderBinaryHeader);
+    hdr.entry_point_length = static_cast<u32>(std::strlen(entry_point));
+    hdr.source_offset = hdr.entry_point_offset + hdr.entry_point_length;
+    hdr.source_length = static_cast<u32>(source.size());
+
+    out_binary->resize(sizeof(hdr) + hdr.entry_point_length + hdr.source_length);
+    std::memcpy(out_binary->data(), &hdr, sizeof(hdr));
+    std::memcpy(&out_binary->data()[hdr.entry_point_offset], entry_point, hdr.entry_point_length);
+    std::memcpy(&out_binary->data()[hdr.source_offset], source.data(), hdr.source_length);
   }
 
-  return CreateShaderFromMSL(stage, msl, "main0", error);
+  return CreateShaderFromMSL(stage, source, entry_point, error);
 }
 
 MetalPipeline::MetalPipeline(id<MTLRenderPipelineState> pipeline, id<MTLDepthStencilState> depth, MTLCullMode cull_mode,