Block is already a merge block for another header #3441
Description
The loop unrolling pass has recently been added to performance passes. We have received a couple of reports saying this has caused the following error to occur for some HLSL shaders compiled with DXC. It seems to be a corner case.
Here is a sample shader:
RWStructuredBuffer<float> output;
struct Data {
uint flags;
float4 extended; // Must be a vector -- scalar float does not repro
};
// This function looks completely pointless, but without it there is no repro
// (obviously it does something in the actual shader!)
void emptyFunction(inout float4 param) {}
bool boolFunction(Data data) {
// This must be a if-else, conditional expression, cast or just if (without else) does not repro
if (data.flags)
return true;
else
return false;
}
Data getData() {
Data data;
emptyFunction(data.extended);
if (boolFunction(data)) {
// This must be a single component, setting the whole vector does not repro
data.extended.x = 0;
}
return data;
}
// Must be >= 2, 1 does not repro
#define LOOP_COUNT 2
[numthreads(64, 1, 1)]
void main() {
// Must be unroll, [loop] does not repro
[unroll]
for(uint i = 0; i < LOOP_COUNT; ++i)
output[0] = getData().extended.x;
}If optimizations are disabled, we get the following SPIR-V:
; SPIR-V
; Version: 1.0
; Generator: Google spiregg; 0
; Bound: 78
; Schema: 0
OpCapability Shader
OpMemoryModel Logical GLSL450
OpEntryPoint GLCompute %main "main"
OpExecutionMode %main LocalSize 64 1 1
OpSource HLSL 600
OpName %type_RWStructuredBuffer_float "type.RWStructuredBuffer.float"
OpName %output "output"
OpName %type_ACSBuffer_counter "type.ACSBuffer.counter"
OpMemberName %type_ACSBuffer_counter 0 "counter"
OpName %counter_var_output "counter.var.output"
OpName %main "main"
OpName %src_main "src.main"
OpName %bb_entry "bb.entry"
OpName %i "i"
OpName %Data "Data"
OpMemberName %Data 0 "flags"
OpMemberName %Data 1 "extended"
OpName %temp_var_Data "temp.var.Data"
OpName %for_check "for.check"
OpName %for_body "for.body"
OpName %for_continue "for.continue"
OpName %for_merge "for.merge"
OpName %getData "getData"
OpName %bb_entry_0 "bb.entry"
OpName %data "data"
OpName %param_var_data "param.var.data"
OpName %if_true "if.true"
OpName %if_merge "if.merge"
OpName %emptyFunction "emptyFunction"
OpName %param "param"
OpName %bb_entry_1 "bb.entry"
OpName %boolFunction "boolFunction"
OpName %data_0 "data"
OpName %bb_entry_2 "bb.entry"
OpName %if_true_0 "if.true"
OpName %if_false "if.false"
OpName %if_merge_0 "if.merge"
OpDecorate %output DescriptorSet 0
OpDecorate %output Binding 0
OpDecorate %counter_var_output DescriptorSet 0
OpDecorate %counter_var_output Binding 1
OpDecorate %_runtimearr_float ArrayStride 4
OpMemberDecorate %type_RWStructuredBuffer_float 0 Offset 0
OpDecorate %type_RWStructuredBuffer_float BufferBlock
OpMemberDecorate %type_ACSBuffer_counter 0 Offset 0
OpDecorate %type_ACSBuffer_counter BufferBlock
%uint = OpTypeInt 32 0
%uint_0 = OpConstant %uint 0
%uint_2 = OpConstant %uint 2
%int = OpTypeInt 32 1
%int_1 = OpConstant %int 1
%int_0 = OpConstant %int 0
%uint_1 = OpConstant %uint 1
%float = OpTypeFloat 32
%float_0 = OpConstant %float 0
%bool = OpTypeBool
%true = OpConstantTrue %bool
%false = OpConstantFalse %bool
%16 = OpConstantNull %bool
%_runtimearr_float = OpTypeRuntimeArray %float
%type_RWStructuredBuffer_float = OpTypeStruct %_runtimearr_float
%_ptr_Uniform_type_RWStructuredBuffer_float = OpTypePointer Uniform %type_RWStructuredBuffer_float
%type_ACSBuffer_counter = OpTypeStruct %int
%_ptr_Uniform_type_ACSBuffer_counter = OpTypePointer Uniform %type_ACSBuffer_counter
%void = OpTypeVoid
%23 = OpTypeFunction %void
%_ptr_Function_uint = OpTypePointer Function %uint
%v4float = OpTypeVector %float 4
%Data = OpTypeStruct %uint %v4float
%_ptr_Function_Data = OpTypePointer Function %Data
%_ptr_Function_v4float = OpTypePointer Function %v4float
%_ptr_Uniform_float = OpTypePointer Uniform %float
%50 = OpTypeFunction %Data
%_ptr_Function_float = OpTypePointer Function %float
%66 = OpTypeFunction %void %_ptr_Function_v4float
%69 = OpTypeFunction %bool %_ptr_Function_Data
%output = OpVariable %_ptr_Uniform_type_RWStructuredBuffer_float Uniform
%counter_var_output = OpVariable %_ptr_Uniform_type_ACSBuffer_counter Uniform
%main = OpFunction %void None %23
%24 = OpLabel
%25 = OpFunctionCall %void %src_main
OpReturn
OpFunctionEnd
%src_main = OpFunction %void None %23
%bb_entry = OpLabel
%i = OpVariable %_ptr_Function_uint Function
%temp_var_Data = OpVariable %_ptr_Function_Data Function
OpStore %i %uint_0
OpBranch %for_check
%for_check = OpLabel
%35 = OpLoad %uint %i
%36 = OpULessThan %bool %35 %uint_2
OpLoopMerge %for_merge %for_continue Unroll
OpBranchConditional %36 %for_body %for_merge
%for_body = OpLabel
%40 = OpFunctionCall %Data %getData
OpStore %temp_var_Data %40
%43 = OpAccessChain %_ptr_Function_v4float %temp_var_Data %int_1
%44 = OpLoad %v4float %43
%45 = OpCompositeExtract %float %44 0
%47 = OpAccessChain %_ptr_Uniform_float %output %int_0 %uint_0
OpStore %47 %45
OpBranch %for_continue
%for_continue = OpLabel
%48 = OpLoad %uint %i
%49 = OpIAdd %uint %48 %uint_1
OpStore %i %49
OpBranch %for_check
%for_merge = OpLabel
OpReturn
OpFunctionEnd
%getData = OpFunction %Data None %50
%bb_entry_0 = OpLabel
%data = OpVariable %_ptr_Function_Data Function
%param_var_data = OpVariable %_ptr_Function_Data Function
%54 = OpAccessChain %_ptr_Function_v4float %data %int_1
%55 = OpFunctionCall %void %emptyFunction %54
%57 = OpLoad %Data %data
OpStore %param_var_data %57
%58 = OpFunctionCall %bool %boolFunction %param_var_data
OpSelectionMerge %if_merge None
OpBranchConditional %58 %if_true %if_merge
%if_true = OpLabel
%62 = OpAccessChain %_ptr_Function_v4float %data %int_1
%64 = OpAccessChain %_ptr_Function_float %62 %int_0
OpStore %64 %float_0
OpBranch %if_merge
%if_merge = OpLabel
%65 = OpLoad %Data %data
OpReturnValue %65
OpFunctionEnd
%emptyFunction = OpFunction %void None %66
%param = OpFunctionParameter %_ptr_Function_v4float
%bb_entry_1 = OpLabel
OpReturn
OpFunctionEnd
%boolFunction = OpFunction %bool None %69
%data_0 = OpFunctionParameter %_ptr_Function_Data
%bb_entry_2 = OpLabel
%72 = OpAccessChain %_ptr_Function_uint %data_0 %int_0
%73 = OpLoad %uint %72
%74 = OpINotEqual %bool %73 %uint_0
OpSelectionMerge %if_merge_0 None
OpBranchConditional %74 %if_true_0 %if_false
%if_true_0 = OpLabel
OpReturnValue %true
%if_false = OpLabel
OpReturnValue %false
%if_merge_0 = OpLabel
OpReturnValue %16
OpFunctionEnd
And the default compile (with optimizations enabled) results in the following error:
Block 26[%26] is already a merge block for another header
And the default compile (with optimizations enabled) and spirv-val silenced, we get:
; SPIR-V
; Version: 1.0
; Generator: Google spiregg; 0
; Bound: 46
; Schema: 0
OpCapability Shader
OpMemoryModel Logical GLSL450
OpEntryPoint GLCompute %main "main"
OpExecutionMode %main LocalSize 64 1 1
OpSource HLSL 600
OpName %type_RWStructuredBuffer_float "type.RWStructuredBuffer.float"
OpName %output "output"
OpName %main "main"
OpDecorate %output DescriptorSet 0
OpDecorate %output Binding 0
OpDecorate %_runtimearr_float ArrayStride 4
OpMemberDecorate %type_RWStructuredBuffer_float 0 Offset 0
OpDecorate %type_RWStructuredBuffer_float BufferBlock
%uint = OpTypeInt 32 0
%uint_0 = OpConstant %uint 0
%int = OpTypeInt 32 1
%int_0 = OpConstant %int 0
%float = OpTypeFloat 32
%float_0 = OpConstant %float 0
%bool = OpTypeBool
%true = OpConstantTrue %bool
%false = OpConstantFalse %bool
%_runtimearr_float = OpTypeRuntimeArray %float
%type_RWStructuredBuffer_float = OpTypeStruct %_runtimearr_float
%_ptr_Uniform_type_RWStructuredBuffer_float = OpTypePointer Uniform %type_RWStructuredBuffer_float
%void = OpTypeVoid
%16 = OpTypeFunction %void
%v4float = OpTypeVector %float 4
%_ptr_Uniform_float = OpTypePointer Uniform %float
%output = OpVariable %_ptr_Uniform_type_RWStructuredBuffer_float Uniform
%19 = OpUndef %uint
%20 = OpUndef %v4float
%21 = OpUndef %v4float
%main = OpFunction %void None %16
%22 = OpLabel
OpSelectionMerge %23 None
OpSwitch %uint_0 %24
%24 = OpLabel
%25 = OpINotEqual %bool %19 %uint_0
OpSelectionMerge %26 None
OpBranchConditional %25 %27 %28
%28 = OpLabel
OpBranch %23
%27 = OpLabel
OpBranch %23
%26 = OpLabel
OpUnreachable
%23 = OpLabel
%29 = OpPhi %bool %false %28 %true %27
OpSelectionMerge %30 None
OpBranchConditional %29 %31 %30
%31 = OpLabel
%32 = OpCompositeInsert %v4float %float_0 %21 0
OpBranch %30
%30 = OpLabel
%33 = OpPhi %v4float %20 %23 %32 %31
%34 = OpCompositeExtract %float %33 0
%35 = OpAccessChain %_ptr_Uniform_float %output %int_0 %uint_0
OpStore %35 %34
OpSelectionMerge %36 None
OpSwitch %uint_0 %37
%37 = OpLabel
OpSelectionMerge %26 None
OpBranchConditional %25 %38 %39
%39 = OpLabel
OpBranch %36
%38 = OpLabel
OpBranch %36
%36 = OpLabel
%40 = OpPhi %bool %false %39 %true %38
OpSelectionMerge %41 None
OpBranchConditional %40 %42 %41
%42 = OpLabel
%43 = OpCompositeInsert %v4float %float_0 %21 0
OpBranch %41
%41 = OpLabel
%44 = OpPhi %v4float %33 %36 %43 %42
%45 = OpCompositeExtract %float %44 0
OpStore %35 %45
OpReturn
OpFunctionEnd
where in fact %26 is used a merge block of two different OpSelectionMerge instructions.