diff --git a/GPU/GLES/DisplayListInterpreter.cpp b/GPU/GLES/DisplayListInterpreter.cpp
index cea3838d5b..d9f10db376 100644
--- a/GPU/GLES/DisplayListInterpreter.cpp
+++ b/GPU/GLES/DisplayListInterpreter.cpp
@@ -921,7 +921,7 @@ void GLES_GPU::ExecuteOp(u32 op, u32 diff) {
 				shaderManager_->DirtyUniform(DIRTY_WORLDMATRIX);
 			}
 			num++;
-			gstate.worldmtxnum = (gstate.worldmtxnum & 0xFF000000) | (num & 0xF);
+			gstate.worldmtxnum = (GE_CMD_WORLDMATRIXNUMBER << 24) | (num & 0xF);
 		}
 		break;
 
@@ -939,7 +939,7 @@ void GLES_GPU::ExecuteOp(u32 op, u32 diff) {
 				shaderManager_->DirtyUniform(DIRTY_VIEWMATRIX);
 			}
 			num++;
-			gstate.viewmtxnum = (gstate.viewmtxnum & 0xFF000000) | (num & 0xF);
+			gstate.viewmtxnum = (GE_CMD_VIEWMATRIXNUMBER << 24) | (num & 0xF);
 		}
 		break;
 
@@ -957,7 +957,7 @@ void GLES_GPU::ExecuteOp(u32 op, u32 diff) {
 				shaderManager_->DirtyUniform(DIRTY_PROJMATRIX | DIRTY_PROJTHROUGHMATRIX);
 			}
 			num++;
-			gstate.projmtxnum = (gstate.projmtxnum & 0xFF000000) | (num & 0xF);
+			gstate.projmtxnum = (GE_CMD_PROJMATRIXNUMBER << 24) | (num & 0xF);
 		}
 		break;
 
@@ -975,7 +975,7 @@ void GLES_GPU::ExecuteOp(u32 op, u32 diff) {
 				shaderManager_->DirtyUniform(DIRTY_TEXMATRIX);
 			}
 			num++;
-			gstate.texmtxnum = (gstate.texmtxnum & 0xFF000000) | (num & 0xF);
+			gstate.texmtxnum = (GE_CMD_TGENMATRIXNUMBER << 24) | (num & 0xF);
 		}
 		break;
 
@@ -993,7 +993,7 @@ void GLES_GPU::ExecuteOp(u32 op, u32 diff) {
 				shaderManager_->DirtyUniform(DIRTY_BONEMATRIX0 << (num / 12));
 			}
 			num++;
-			gstate.boneMatrixNumber = (gstate.boneMatrixNumber & 0xFF000000) | (num & 0x7F);
+			gstate.boneMatrixNumber = (GE_CMD_BONEMATRIXNUMBER << 24) | (num & 0x7F);
 		}
 		break;
 
@@ -1120,11 +1120,6 @@ void GLES_GPU::ClearCacheNextFrame() {
 	textureCache_.ClearNextFrame();
 }
 
-
-inline void GLES_GPU::Flush() {
-	transformDraw_.Flush();
-}
-
 void GLES_GPU::Resized() {
 	framebufferManager_.Resized();
 }
diff --git a/GPU/GLES/DisplayListInterpreter.h b/GPU/GLES/DisplayListInterpreter.h
index aa97f25fd0..76a4969480 100644
--- a/GPU/GLES/DisplayListInterpreter.h
+++ b/GPU/GLES/DisplayListInterpreter.h
@@ -69,7 +69,9 @@ protected:
 	virtual void ProcessEvent(GPUEvent ev);
 
 private:
-	void Flush();
+	void Flush() {
+		transformDraw_.Flush();
+	}
 	void DoBlockTransfer();
 	void ApplyDrawState(int prim);
 	void CheckFlushOp(int cmd, u32 diff);
diff --git a/GPU/GLES/TransformPipeline.cpp b/GPU/GLES/TransformPipeline.cpp
index c087af9f74..a5a5e1cd0a 100644
--- a/GPU/GLES/TransformPipeline.cpp
+++ b/GPU/GLES/TransformPipeline.cpp
@@ -1109,10 +1109,7 @@ VertexArrayInfo::~VertexArrayInfo() {
 		glDeleteBuffers(1, &ebo);
 }
 
-void TransformDrawEngine::Flush() {
-	if (!numDrawCalls)
-		return;
-
+void TransformDrawEngine::DoFlush() {
 	gpuStats.numFlushes++;
 	
 	gpuStats.numTrackedVertexArrays = (int)vai_.size();
diff --git a/GPU/GLES/TransformPipeline.h b/GPU/GLES/TransformPipeline.h
index 20391c2163..98474748e1 100644
--- a/GPU/GLES/TransformPipeline.h
+++ b/GPU/GLES/TransformPipeline.h
@@ -100,7 +100,6 @@ public:
 	void DrawBezier(int ucount, int vcount);
 	void DrawSpline(int ucount, int vcount, int utype, int vtype);
 	void DecodeVerts();
-	void Flush();
 	void SetShaderManager(ShaderManager *shaderManager) {
 		shaderManager_ = shaderManager;
 	}
@@ -122,7 +121,15 @@ public:
 	// This requires a SetupVertexDecoder call first.
 	int EstimatePerVertexCost();
 
+	// So that this can be inlined
+	void Flush() {
+		if (!numDrawCalls)
+			return;
+		DoFlush();
+	}
+
 private:
+	void DoFlush();
 	void SoftwareTransformAndDraw(int prim, u8 *decoded, LinkedShader *program, int vertexCount, u32 vertexType, void *inds, int indexType, const DecVtxFormat &decVtxFormat, int maxIndex);
 	void ApplyDrawState(int prim);
 	bool IsReallyAClear(int numVerts) const;