Merge pull request #1043 from KhronosGroup/fix-1042

Fix declaration of loop variables with a Phi helper copy.

reference/opt/shaders/asm/comp/phi-temporary-copy-loop-variable.asm.comp

@@ -0,0 +1,25 @@
+#version 450
+layout(local_size_x = 8, local_size_y = 8, local_size_z = 1) in;
+
+layout(binding = 1, rgba32f) uniform writeonly image2D outImageTexture;
+
+void main()
+{
+    int _30;
+    _30 = 7;
+    int _27_copy;
+    for (int _27 = 7; _27 >= 0; _27_copy = _27, _27--, _30 = _27_copy)
+    {
+        if (5.0 > float(_27))
+        {
+            break;
+        }
+        else
+        {
+            continue;
+        }
+        continue;
+    }
+    imageStore(outImageTexture, ivec2(gl_GlobalInvocationID.xy), vec4(float(_30 - 1), float(_30), 1.0, 1.0));
+}
+

reference/shaders/asm/comp/phi-temporary-copy-loop-variable.asm.comp

@@ -0,0 +1,25 @@
+#version 450
+layout(local_size_x = 8, local_size_y = 8, local_size_z = 1) in;
+
+layout(binding = 1, rgba32f) uniform writeonly image2D outImageTexture;
+
+void main()
+{
+    int _30;
+    _30 = 7;
+    int _27_copy;
+    for (int _27 = 7; _27 >= 0; _27_copy = _27, _27--, _30 = _27_copy)
+    {
+        if (5.0 > float(_27))
+        {
+            break;
+        }
+        else
+        {
+            continue;
+        }
+        continue;
+    }
+    imageStore(outImageTexture, ivec2(gl_GlobalInvocationID.xy), vec4(float(_30 - 1), float(_30), 1.0, 1.0));
+}
+

shaders/asm/comp/phi-temporary-copy-loop-variable.asm.comp

@@ -0,0 +1,68 @@
+; SPIR-V
+; Version: 1.0
+; Generator: Google spiregg; 0
+; Bound: 42
+; Schema: 0
+               OpCapability Shader
+               OpMemoryModel Logical GLSL450
+               OpEntryPoint GLCompute %cs_test "main" %gl_GlobalInvocationID %gl_LocalInvocationIndex
+               OpExecutionMode %cs_test LocalSize 8 8 1
+               OpSource HLSL 600
+               OpName %type_2d_image "type.2d.image"
+               OpName %outImageTexture "outImageTexture"
+               OpName %cs_test "cs_test"
+               OpDecorate %gl_GlobalInvocationID BuiltIn GlobalInvocationId
+               OpDecorate %gl_LocalInvocationIndex BuiltIn LocalInvocationIndex
+               OpDecorate %outImageTexture DescriptorSet 0
+               OpDecorate %outImageTexture Binding 1
+      %float = OpTypeFloat 32
+    %float_5 = OpConstant %float 5
+    %float_1 = OpConstant %float 1
+        %int = OpTypeInt 32 1
+      %int_7 = OpConstant %int 7
+      %int_0 = OpConstant %int 0
+      %int_1 = OpConstant %int 1
+%type_2d_image = OpTypeImage %float 2D 2 0 0 2 Rgba32f
+%_ptr_UniformConstant_type_2d_image = OpTypePointer UniformConstant %type_2d_image
+       %uint = OpTypeInt 32 0
+     %v3uint = OpTypeVector %uint 3
+%_ptr_Input_v3uint = OpTypePointer Input %v3uint
+%_ptr_Input_uint = OpTypePointer Input %uint
+       %void = OpTypeVoid
+         %19 = OpTypeFunction %void
+     %v2uint = OpTypeVector %uint 2
+    %v4float = OpTypeVector %float 4
+       %bool = OpTypeBool
+%outImageTexture = OpVariable %_ptr_UniformConstant_type_2d_image UniformConstant
+%gl_GlobalInvocationID = OpVariable %_ptr_Input_v3uint Input
+%gl_LocalInvocationIndex = OpVariable %_ptr_Input_uint Input
+    %cs_test = OpFunction %void None %19
+         %23 = OpLabel
+         %24 = OpLoad %v3uint %gl_GlobalInvocationID
+         %25 = OpVectorShuffle %v2uint %24 %24 0 1
+               OpBranch %26
+         %26 = OpLabel
+         %27 = OpPhi %int %int_7 %23 %28 %29
+         %30 = OpPhi %int %int_7 %23 %27 %29
+         %31 = OpSGreaterThanEqual %bool %27 %int_0
+               OpLoopMerge %32 %29 None
+               OpBranchConditional %31 %33 %32
+         %33 = OpLabel
+         %34 = OpConvertSToF %float %27
+         %35 = OpFOrdGreaterThan %bool %float_5 %34
+               OpSelectionMerge %29 None
+               OpBranchConditional %35 %36 %29
+         %36 = OpLabel
+               OpBranch %32
+         %29 = OpLabel
+         %28 = OpISub %int %27 %int_1
+               OpBranch %26
+         %32 = OpLabel
+         %37 = OpISub %int %30 %int_1
+         %38 = OpConvertSToF %float %37
+         %39 = OpConvertSToF %float %30
+         %40 = OpCompositeConstruct %v4float %38 %39 %float_1 %float_1
+         %41 = OpLoad %type_2d_image %outImageTexture
+               OpImageWrite %41 %25 %40 None
+               OpReturn
+               OpFunctionEnd

spirv_glsl.cpp

@@ -7107,18 +7107,23 @@ string CompilerGLSL::variable_decl_function_local(SPIRVariable &var)
 	return expr;
 }
 
+void CompilerGLSL::emit_variable_temporary_copies(const SPIRVariable &var)
+{
+	if (var.allocate_temporary_copy)
+	{
+		auto &type = get<SPIRType>(var.basetype);
+		auto &flags = get_decoration_bitset(var.self);
+		statement(flags_to_qualifiers_glsl(type, flags), variable_decl(type, join("_", var.self, "_copy")), ";");
+	}
+}
+
 void CompilerGLSL::flush_variable_declaration(uint32_t id)
 {
 	auto *var = maybe_get<SPIRVariable>(id);
 	if (var && var->deferred_declaration)
 	{
 		statement(variable_decl_function_local(*var), ";");
-		if (var->allocate_temporary_copy)
+		emit_variable_temporary_copies(*var);
-		{
-			auto &type = get<SPIRType>(var->basetype);
-			auto &flags = ir.meta[id].decoration.decoration_flags;
-			statement(flags_to_qualifiers_glsl(type, flags), variable_decl(type, join("_", id, "_copy")), ";");
-		}
 		var->deferred_declaration = false;
 	}
 }
@@ -11325,8 +11330,13 @@ void CompilerGLSL::emit_block_chain(SPIRBlock &block)
 		continue_type = continue_block_type(get<SPIRBlock>(block.continue_block));
 
 	// If we have loop variables, stop masking out access to the variable now.
-	for (auto var : block.loop_variables)
+	for (auto var_id : block.loop_variables)
-		get<SPIRVariable>(var).loop_variable_enable = true;
+	{
+		auto &var = get<SPIRVariable>(var_id);
+		var.loop_variable_enable = true;
+		// We're not going to declare the variable directly, so emit a copy here.
+		emit_variable_temporary_copies(var);
+	}
 
 	// Remember deferred declaration state. We will restore it before returning.
 	SmallVector<bool, 64> rearm_dominated_variables(block.dominated_variables.size());

spirv_glsl.hpp

@@ -432,6 +432,7 @@ protected:
 	bool flush_phi_required(uint32_t from, uint32_t to);
 	void flush_variable_declaration(uint32_t id);
 	void flush_undeclared_variables(SPIRBlock &block);
+	void emit_variable_temporary_copies(const SPIRVariable &var);
 
 	bool should_dereference(uint32_t id);
 	bool should_forward(uint32_t id);