Fix a lot of redundant code when loading flattened composites.

reference/opt/shaders/flatten/copy.flatten.vert

@@ -18,8 +18,9 @@ void main()
     vColor = vec4(0.0);
     for (int _96 = 0; _96 < 4; )
     {
-        vec3 _68 = aVertex.xyz - Light(UBO[_96 * 2 + 4].xyz, UBO[_96 * 2 + 4].w, UBO[_96 * 2 + 5]).Position;
-        vColor += ((UBO[_96 * 2 + 5] * clamp(1.0 - (length(_68) / Light(UBO[_96 * 2 + 4].xyz, UBO[_96 * 2 + 4].w, UBO[_96 * 2 + 5]).Radius), 0.0, 1.0)) * dot(aNormal, normalize(_68)));
+        Light _52 = Light(UBO[_96 * 2 + 4].xyz, UBO[_96 * 2 + 4].w, UBO[_96 * 2 + 5]);
+        vec3 _68 = aVertex.xyz - _52.Position;
+        vColor += ((UBO[_96 * 2 + 5] * clamp(1.0 - (length(_68) / _52.Radius), 0.0, 1.0)) * dot(aNormal, normalize(_68)));
         _96++;
         continue;
     }

reference/opt/shaders/flatten/matrix-conversion.flatten.frag

@@ -0,0 +1,14 @@
+#version 310 es
+precision mediump float;
+precision highp int;
+
+uniform vec4 UBO[4];
+layout(location = 0) out vec3 FragColor;
+layout(location = 0) flat in vec3 vNormal;
+
+void main()
+{
+    mat4 _19 = mat4(UBO[0], UBO[1], UBO[2], UBO[3]);
+    FragColor = mat3(vec3(_19[0].x, _19[0].y, _19[0].z), vec3(_19[1].x, _19[1].y, _19[1].z), vec3(_19[2].x, _19[2].y, _19[2].z)) * vNormal;
+}
+

reference/opt/shaders/flatten/struct.rowmajor.flatten.vert

@@ -14,7 +14,8 @@ layout(location = 1) out vec3 V1;
 
 void main()
 {
-    V0 = v0 * Foo(transpose(mat4x3(UBO[0].xyz, UBO[1].xyz, UBO[2].xyz, UBO[3].xyz)), transpose(mat4x3(UBO[4].xyz, UBO[5].xyz, UBO[6].xyz, UBO[7].xyz))).MVP0;
-    V1 = v1 * Foo(transpose(mat4x3(UBO[0].xyz, UBO[1].xyz, UBO[2].xyz, UBO[3].xyz)), transpose(mat4x3(UBO[4].xyz, UBO[5].xyz, UBO[6].xyz, UBO[7].xyz))).MVP1;
+    Foo _20 = Foo(transpose(mat4x3(UBO[0].xyz, UBO[1].xyz, UBO[2].xyz, UBO[3].xyz)), transpose(mat4x3(UBO[4].xyz, UBO[5].xyz, UBO[6].xyz, UBO[7].xyz)));
+    V0 = v0 * _20.MVP0;
+    V1 = v1 * _20.MVP1;
 }
 

reference/shaders/flatten/copy.flatten.vert

@@ -18,10 +18,11 @@ void main()
     vColor = vec4(0.0);
     for (int i = 0; i < 4; i++)
     {
+        Light _52 = Light(UBO[i * 2 + 4].xyz, UBO[i * 2 + 4].w, UBO[i * 2 + 5]);
         Light light;
-        light.Position = Light(UBO[i * 2 + 4].xyz, UBO[i * 2 + 4].w, UBO[i * 2 + 5]).Position;
-        light.Radius = Light(UBO[i * 2 + 4].xyz, UBO[i * 2 + 4].w, UBO[i * 2 + 5]).Radius;
-        light.Color = Light(UBO[i * 2 + 4].xyz, UBO[i * 2 + 4].w, UBO[i * 2 + 5]).Color;
+        light.Position = _52.Position;
+        light.Radius = _52.Radius;
+        light.Color = _52.Color;
         vec3 L = aVertex.xyz - light.Position;
         vColor += ((UBO[i * 2 + 5] * clamp(1.0 - (length(L) / light.Radius), 0.0, 1.0)) * dot(aNormal, normalize(L)));
     }

reference/shaders/flatten/matrix-conversion.flatten.frag

@@ -0,0 +1,14 @@
+#version 310 es
+precision mediump float;
+precision highp int;
+
+uniform vec4 UBO[4];
+layout(location = 0) out vec3 FragColor;
+layout(location = 0) flat in vec3 vNormal;
+
+void main()
+{
+    mat4 _19 = mat4(UBO[0], UBO[1], UBO[2], UBO[3]);
+    FragColor = mat3(vec3(_19[0].x, _19[0].y, _19[0].z), vec3(_19[1].x, _19[1].y, _19[1].z), vec3(_19[2].x, _19[2].y, _19[2].z)) * vNormal;
+}
+

reference/shaders/flatten/struct.rowmajor.flatten.vert

@@ -14,9 +14,10 @@ layout(location = 1) out vec3 V1;
 
 void main()
 {
+    Foo _20 = Foo(transpose(mat4x3(UBO[0].xyz, UBO[1].xyz, UBO[2].xyz, UBO[3].xyz)), transpose(mat4x3(UBO[4].xyz, UBO[5].xyz, UBO[6].xyz, UBO[7].xyz)));
     Foo f;
-    f.MVP0 = Foo(transpose(mat4x3(UBO[0].xyz, UBO[1].xyz, UBO[2].xyz, UBO[3].xyz)), transpose(mat4x3(UBO[4].xyz, UBO[5].xyz, UBO[6].xyz, UBO[7].xyz))).MVP0;
-    f.MVP1 = Foo(transpose(mat4x3(UBO[0].xyz, UBO[1].xyz, UBO[2].xyz, UBO[3].xyz)), transpose(mat4x3(UBO[4].xyz, UBO[5].xyz, UBO[6].xyz, UBO[7].xyz))).MVP1;
+    f.MVP0 = _20.MVP0;
+    f.MVP1 = _20.MVP1;
     vec3 a = v0 * f.MVP0;
     vec3 b = v1 * f.MVP1;
     V0 = a;

shaders/flatten/matrix-conversion.flatten.frag

@@ -0,0 +1,14 @@
+#version 310 es
+precision mediump float;
+layout(location = 0) out vec3 FragColor;
+layout(location = 0) flat in vec3 vNormal;
+
+layout(binding = 0, std140) uniform UBO
+{
+	mat4 m;
+};
+
+void main()
+{
+	FragColor = mat3(m) * vNormal;
+}

spirv_glsl.cpp

@@ -6147,8 +6147,14 @@ void CompilerGLSL::emit_instruction(const Instruction &instruction)
 		if (ptr_expression)
 			ptr_expression->need_transpose = old_need_transpose;
 
-		// Suppress usage tracking since using same expression multiple times does not imply any extra work.
-		auto &e = emit_op(result_type, id, expr, forward, true);
+		// By default, suppress usage tracking since using same expressio multiple times does not imply any extra work.
+		// However, if we try to load a complex, composite object from a flattened buffer,
+		// we should avoid emitting the same code over and over and lower the result to a temporary.
+		auto &type = get<SPIRType>(result_type);
+		bool usage_tracking = ptr_expression && flattened_buffer_blocks.count(ptr_expression->loaded_from) != 0 &&
+		                      (type.basetype == SPIRType::Struct || (type.columns > 1));
+
+		auto &e = emit_op(result_type, id, expr, forward, !usage_tracking);
 		e.need_transpose = need_transpose;
 		register_read(id, ptr, forward);
 
@@ -6168,8 +6174,10 @@ void CompilerGLSL::emit_instruction(const Instruction &instruction)
 
 		// If the base is immutable, the access chain pointer must also be.
 		// If an expression is mutable and forwardable, we speculate that it is immutable.
-		bool need_transpose, result_is_packed;
+		bool need_transpose = false;
+		bool result_is_packed = false;
 		auto e = access_chain(ops[2], &ops[3], length - 3, get<SPIRType>(ops[0]), &need_transpose, &result_is_packed);
+
 		auto &expr = set<SPIRExpression>(ops[1], move(e), ops[0], should_forward(ops[2]));
 
 		auto *backing_variable = maybe_get_backing_variable(ops[2]);