Merge pull request #806 from KhronosGroup/fix-793
Fix case where OpPhi is used to swap values.
This commit is contained in:
Hans-Kristian Arntzen
GitHub
commit
b30d21bafd
39
reference/opt/shaders/asm/comp/op-phi-swap.asm.comp
Normal file
39
reference/opt/shaders/asm/comp/op-phi-swap.asm.comp
Normal file
@ -0,0 +1,39 @@
|
||||
#version 450
|
||||
layout(local_size_x = 1, local_size_y = 1, local_size_z = 1) in;
|
||||
|
||||
layout(binding = 0, std430) buffer _3_4
|
||||
{
|
||||
float _m0[];
|
||||
} _4;
|
||||
|
||||
layout(binding = 1, std430) buffer _3_5
|
||||
{
|
||||
float _m0[];
|
||||
} _5;
|
||||
|
||||
void main()
|
||||
{
|
||||
bool _34;
|
||||
float _35;
|
||||
float _35_copy;
|
||||
float _36;
|
||||
_34 = true;
|
||||
_35 = _4._m0[gl_GlobalInvocationID.x];
|
||||
_36 = 8.5;
|
||||
for (;;)
|
||||
{
|
||||
if (_34)
|
||||
{
|
||||
_34 = false;
|
||||
_35_copy = _35;
|
||||
_35 = _36;
|
||||
_36 = _35_copy;
|
||||
}
|
||||
else
|
||||
{
|
||||
break;
|
||||
}
|
||||
}
|
||||
_5._m0[gl_GlobalInvocationID.x] = _35 - _36;
|
||||
}
|
||||
|
||||
@ -0,0 +1,29 @@
|
||||
#version 450
|
||||
|
||||
layout(binding = 0, std140) uniform UBO
|
||||
{
|
||||
int uCount;
|
||||
int uJ;
|
||||
int uK;
|
||||
} _5;
|
||||
|
||||
layout(location = 0) out float FragColor;
|
||||
|
||||
void main()
|
||||
{
|
||||
int _23;
|
||||
int _23_copy;
|
||||
int _24;
|
||||
_23 = _5.uK;
|
||||
_24 = _5.uJ;
|
||||
for (int _26 = 0; _26 < _5.uCount; )
|
||||
{
|
||||
_23_copy = _23;
|
||||
_23 = _24;
|
||||
_24 = _23_copy;
|
||||
_26++;
|
||||
continue;
|
||||
}
|
||||
FragColor = float(_24 - _23) * float(_5.uJ * _5.uK);
|
||||
}
|
||||
|
||||
40
reference/shaders/asm/comp/op-phi-swap.asm.comp
Normal file
40
reference/shaders/asm/comp/op-phi-swap.asm.comp
Normal file
@ -0,0 +1,40 @@
|
||||
#version 450
|
||||
layout(local_size_x = 1, local_size_y = 1, local_size_z = 1) in;
|
||||
|
||||
layout(binding = 0, std430) buffer _3_4
|
||||
{
|
||||
float _m0[];
|
||||
} _4;
|
||||
|
||||
layout(binding = 1, std430) buffer _3_5
|
||||
{
|
||||
float _m0[];
|
||||
} _5;
|
||||
|
||||
void main()
|
||||
{
|
||||
float _26 = 8.5;
|
||||
bool _34;
|
||||
float _35;
|
||||
float _35_copy;
|
||||
float _36;
|
||||
_34 = true;
|
||||
_35 = _4._m0[gl_GlobalInvocationID.x];
|
||||
_36 = _26;
|
||||
for (;;)
|
||||
{
|
||||
if (_34)
|
||||
{
|
||||
_34 = false;
|
||||
_35_copy = _35;
|
||||
_35 = _36;
|
||||
_36 = _35_copy;
|
||||
}
|
||||
else
|
||||
{
|
||||
break;
|
||||
}
|
||||
}
|
||||
_5._m0[gl_GlobalInvocationID.x] = _35 - _36;
|
||||
}
|
||||
|
||||
@ -0,0 +1,25 @@
|
||||
#version 450
|
||||
|
||||
layout(binding = 0, std140) uniform UBO
|
||||
{
|
||||
int uCount;
|
||||
int uJ;
|
||||
int uK;
|
||||
} _5;
|
||||
|
||||
layout(location = 0) out float FragColor;
|
||||
|
||||
void main()
|
||||
{
|
||||
int _23;
|
||||
int _23_copy;
|
||||
int _24;
|
||||
_23 = _5.uK;
|
||||
_24 = _5.uJ;
|
||||
for (int _26 = 0; _26 < _5.uCount; _23_copy = _23, _23 = _24, _24 = _23_copy, _26++)
|
||||
{
|
||||
continue;
|
||||
}
|
||||
FragColor = float(_24 - _23) * float(_5.uJ * _5.uK);
|
||||
}
|
||||
|
||||
63
shaders/asm/comp/op-phi-swap.asm.comp
Normal file
63
shaders/asm/comp/op-phi-swap.asm.comp
Normal file
@ -0,0 +1,63 @@
|
||||
; SPIR-V
|
||||
; Version: 1.0
|
||||
; Generator: Khronos SPIR-V Tools Assembler; 0
|
||||
; Bound: 39
|
||||
; Schema: 0
|
||||
OpCapability Shader
|
||||
OpMemoryModel Logical GLSL450
|
||||
OpEntryPoint GLCompute %main "main" %gl_GlobalInvocationID
|
||||
OpExecutionMode %main LocalSize 1 1 1
|
||||
OpName %main "main"
|
||||
OpName %gl_GlobalInvocationID "gl_GlobalInvocationID"
|
||||
OpDecorate %gl_GlobalInvocationID BuiltIn GlobalInvocationId
|
||||
OpDecorate %_struct_3 BufferBlock
|
||||
OpDecorate %4 DescriptorSet 0
|
||||
OpDecorate %4 Binding 0
|
||||
OpDecorate %5 DescriptorSet 0
|
||||
OpDecorate %5 Binding 1
|
||||
OpDecorate %_runtimearr_float ArrayStride 4
|
||||
OpMemberDecorate %_struct_3 0 Offset 0
|
||||
%bool = OpTypeBool
|
||||
%void = OpTypeVoid
|
||||
%9 = OpTypeFunction %void
|
||||
%uint = OpTypeInt 32 0
|
||||
%int = OpTypeInt 32 1
|
||||
%float = OpTypeFloat 32
|
||||
%v3uint = OpTypeVector %uint 3
|
||||
%v3float = OpTypeVector %float 3
|
||||
%_ptr_Input_v3uint = OpTypePointer Input %v3uint
|
||||
%_ptr_Uniform_int = OpTypePointer Uniform %int
|
||||
%_ptr_Uniform_float = OpTypePointer Uniform %float
|
||||
%_runtimearr_int = OpTypeRuntimeArray %int
|
||||
%_runtimearr_float = OpTypeRuntimeArray %float
|
||||
%_struct_3 = OpTypeStruct %_runtimearr_float
|
||||
%_ptr_Uniform__struct_3 = OpTypePointer Uniform %_struct_3
|
||||
%4 = OpVariable %_ptr_Uniform__struct_3 Uniform
|
||||
%5 = OpVariable %_ptr_Uniform__struct_3 Uniform
|
||||
%_ptr_Function_float = OpTypePointer Function %float
|
||||
%gl_GlobalInvocationID = OpVariable %_ptr_Input_v3uint Input
|
||||
%true = OpConstantTrue %bool
|
||||
%false = OpConstantFalse %bool
|
||||
%int_0 = OpConstant %int 0
|
||||
%float_8_5 = OpConstant %float 8.5
|
||||
%main = OpFunction %void None %9
|
||||
%25 = OpLabel
|
||||
%26 = OpVariable %_ptr_Function_float Function %float_8_5
|
||||
%27 = OpLoad %v3uint %gl_GlobalInvocationID
|
||||
%28 = OpCompositeExtract %uint %27 0
|
||||
%29 = OpAccessChain %_ptr_Uniform_float %4 %int_0 %28
|
||||
%30 = OpAccessChain %_ptr_Uniform_float %5 %int_0 %28
|
||||
%31 = OpLoad %float %29
|
||||
%32 = OpLoad %float %26
|
||||
OpBranch %33
|
||||
%33 = OpLabel
|
||||
%34 = OpPhi %bool %true %25 %false %33
|
||||
%35 = OpPhi %float %31 %25 %36 %33
|
||||
%36 = OpPhi %float %32 %25 %35 %33
|
||||
OpLoopMerge %37 %33 None
|
||||
OpBranchConditional %34 %33 %37
|
||||
%37 = OpLabel
|
||||
%38 = OpFSub %float %35 %36
|
||||
OpStore %30 %38
|
||||
OpReturn
|
||||
OpFunctionEnd
|
||||
69
shaders/asm/frag/op-phi-swap-continue-block.asm.frag
Normal file
69
shaders/asm/frag/op-phi-swap-continue-block.asm.frag
Normal file
@ -0,0 +1,69 @@
|
||||
; SPIR-V
|
||||
; Version: 1.0
|
||||
; Generator: Khronos Glslang Reference Front End; 7
|
||||
; Bound: 55
|
||||
; Schema: 0
|
||||
OpCapability Shader
|
||||
%1 = OpExtInstImport "GLSL.std.450"
|
||||
OpMemoryModel Logical GLSL450
|
||||
OpEntryPoint Fragment %main "main" %FragColor
|
||||
OpExecutionMode %main OriginUpperLeft
|
||||
OpSource GLSL 450
|
||||
OpName %main "main"
|
||||
OpName %UBO "UBO"
|
||||
OpMemberName %UBO 0 "uCount"
|
||||
OpMemberName %UBO 1 "uJ"
|
||||
OpMemberName %UBO 2 "uK"
|
||||
OpName %_ ""
|
||||
OpName %FragColor "FragColor"
|
||||
OpMemberDecorate %UBO 0 Offset 0
|
||||
OpMemberDecorate %UBO 1 Offset 4
|
||||
OpMemberDecorate %UBO 2 Offset 8
|
||||
OpDecorate %UBO Block
|
||||
OpDecorate %_ DescriptorSet 0
|
||||
OpDecorate %_ Binding 0
|
||||
OpDecorate %FragColor Location 0
|
||||
%void = OpTypeVoid
|
||||
%3 = OpTypeFunction %void
|
||||
%int = OpTypeInt 32 1
|
||||
%UBO = OpTypeStruct %int %int %int
|
||||
%_ptr_Uniform_UBO = OpTypePointer Uniform %UBO
|
||||
%_ = OpVariable %_ptr_Uniform_UBO Uniform
|
||||
%int_1 = OpConstant %int 1
|
||||
%_ptr_Uniform_int = OpTypePointer Uniform %int
|
||||
%int_2 = OpConstant %int 2
|
||||
%int_0 = OpConstant %int 0
|
||||
%bool = OpTypeBool
|
||||
%float = OpTypeFloat 32
|
||||
%_ptr_Output_float = OpTypePointer Output %float
|
||||
%FragColor = OpVariable %_ptr_Output_float Output
|
||||
%main = OpFunction %void None %3
|
||||
%5 = OpLabel
|
||||
%14 = OpAccessChain %_ptr_Uniform_int %_ %int_1
|
||||
%15 = OpLoad %int %14
|
||||
%18 = OpAccessChain %_ptr_Uniform_int %_ %int_2
|
||||
%19 = OpLoad %int %18
|
||||
OpBranch %22
|
||||
%22 = OpLabel
|
||||
%54 = OpPhi %int %19 %5 %53 %23
|
||||
%53 = OpPhi %int %15 %5 %54 %23
|
||||
%52 = OpPhi %int %int_0 %5 %37 %23
|
||||
%28 = OpAccessChain %_ptr_Uniform_int %_ %int_0
|
||||
%29 = OpLoad %int %28
|
||||
%31 = OpSLessThan %bool %52 %29
|
||||
OpLoopMerge %24 %23 None
|
||||
OpBranchConditional %31 %inbetween %24
|
||||
%inbetween = OpLabel
|
||||
OpBranch %23
|
||||
%23 = OpLabel
|
||||
%37 = OpIAdd %int %52 %int_1
|
||||
OpBranch %22
|
||||
%24 = OpLabel
|
||||
%43 = OpISub %int %53 %54
|
||||
%44 = OpConvertSToF %float %43
|
||||
%49 = OpIMul %int %15 %19
|
||||
%50 = OpConvertSToF %float %49
|
||||
%51 = OpFMul %float %44 %50
|
||||
OpStore %FragColor %51
|
||||
OpReturn
|
||||
OpFunctionEnd
|
||||
@ -894,6 +894,10 @@ struct SPIRVariable : IVariant
|
||||
|
||||
bool deferred_declaration = false;
|
||||
bool phi_variable = false;
|
||||
|
||||
// Used to deal with Phi variable flushes. See flush_phi().
|
||||
bool allocate_temporary_copy = false;
|
||||
|
||||
bool remapped_variable = false;
|
||||
uint32_t remapped_components = 0;
|
||||
|
||||
|
||||
@ -1283,7 +1283,7 @@ string CompilerGLSL::layout_for_variable(const SPIRVariable &var)
|
||||
|
||||
auto &dec = ir.meta[var.self].decoration;
|
||||
auto &type = get<SPIRType>(var.basetype);
|
||||
auto flags = dec.decoration_flags;
|
||||
auto &flags = dec.decoration_flags;
|
||||
auto typeflags = ir.meta[type.self].decoration.decoration_flags;
|
||||
|
||||
if (options.vulkan_semantics && var.storage == StorageClassPushConstant)
|
||||
@ -3459,7 +3459,7 @@ string CompilerGLSL::constant_expression_vector(const SPIRConstant &c, uint32_t
|
||||
string CompilerGLSL::declare_temporary(uint32_t result_type, uint32_t result_id)
|
||||
{
|
||||
auto &type = get<SPIRType>(result_type);
|
||||
auto flags = ir.meta[result_id].decoration.decoration_flags;
|
||||
auto &flags = ir.meta[result_id].decoration.decoration_flags;
|
||||
|
||||
// If we're declaring temporaries inside continue blocks,
|
||||
// we must declare the temporary in the loop header so that the continue block can avoid declaring new variables.
|
||||
@ -4568,7 +4568,7 @@ void CompilerGLSL::emit_glsl_op(uint32_t result_type, uint32_t id, uint32_t eop,
|
||||
{
|
||||
forced_temporaries.insert(id);
|
||||
auto &type = get<SPIRType>(result_type);
|
||||
auto flags = ir.meta[id].decoration.decoration_flags;
|
||||
auto &flags = ir.meta[id].decoration.decoration_flags;
|
||||
statement(flags_to_precision_qualifiers_glsl(type, flags), variable_decl(type, to_name(id)), ";");
|
||||
set<SPIRExpression>(id, to_name(id), result_type, true);
|
||||
|
||||
@ -4691,7 +4691,7 @@ void CompilerGLSL::emit_glsl_op(uint32_t result_type, uint32_t id, uint32_t eop,
|
||||
{
|
||||
forced_temporaries.insert(id);
|
||||
auto &type = get<SPIRType>(result_type);
|
||||
auto flags = ir.meta[id].decoration.decoration_flags;
|
||||
auto &flags = ir.meta[id].decoration.decoration_flags;
|
||||
statement(flags_to_precision_qualifiers_glsl(type, flags), variable_decl(type, to_name(id)), ";");
|
||||
set<SPIRExpression>(id, to_name(id), result_type, true);
|
||||
|
||||
@ -6219,6 +6219,13 @@ void CompilerGLSL::flush_variable_declaration(uint32_t id)
|
||||
if (var && var->deferred_declaration)
|
||||
{
|
||||
statement(variable_decl_function_local(*var), ";");
|
||||
if (var->allocate_temporary_copy)
|
||||
{
|
||||
auto &type = get<SPIRType>(var->basetype);
|
||||
auto &flags = ir.meta[id].decoration.decoration_flags;
|
||||
statement(flags_to_precision_qualifiers_glsl(type, flags), variable_decl(type, join("_", id, "_copy")),
|
||||
";");
|
||||
}
|
||||
var->deferred_declaration = false;
|
||||
}
|
||||
}
|
||||
@ -6840,7 +6847,7 @@ void CompilerGLSL::emit_instruction(const Instruction &instruction)
|
||||
// We cannot construct array of arrays because we cannot treat the inputs
|
||||
// as value types. Need to declare the array-of-arrays, and copy in elements one by one.
|
||||
forced_temporaries.insert(id);
|
||||
auto flags = ir.meta[id].decoration.decoration_flags;
|
||||
auto &flags = ir.meta[id].decoration.decoration_flags;
|
||||
statement(flags_to_precision_qualifiers_glsl(out_type, flags), variable_decl(out_type, to_name(id)), ";");
|
||||
set<SPIRExpression>(id, to_name(id), result_type, true);
|
||||
for (uint32_t i = 0; i < length; i++)
|
||||
@ -7225,7 +7232,7 @@ void CompilerGLSL::emit_instruction(const Instruction &instruction)
|
||||
uint32_t op1 = ops[3];
|
||||
forced_temporaries.insert(result_id);
|
||||
auto &type = get<SPIRType>(result_type);
|
||||
auto flags = ir.meta[result_id].decoration.decoration_flags;
|
||||
auto &flags = ir.meta[result_id].decoration.decoration_flags;
|
||||
statement(flags_to_precision_qualifiers_glsl(type, flags), variable_decl(type, to_name(result_id)), ";");
|
||||
set<SPIRExpression>(result_id, to_name(result_id), result_type, true);
|
||||
|
||||
@ -7250,7 +7257,7 @@ void CompilerGLSL::emit_instruction(const Instruction &instruction)
|
||||
uint32_t op1 = ops[3];
|
||||
forced_temporaries.insert(result_id);
|
||||
auto &type = get<SPIRType>(result_type);
|
||||
auto flags = ir.meta[result_id].decoration.decoration_flags;
|
||||
auto &flags = ir.meta[result_id].decoration.decoration_flags;
|
||||
statement(flags_to_precision_qualifiers_glsl(type, flags), variable_decl(type, to_name(result_id)), ";");
|
||||
set<SPIRExpression>(result_id, to_name(result_id), result_type, true);
|
||||
|
||||
@ -8791,7 +8798,7 @@ const char *CompilerGLSL::to_precision_qualifiers_glsl(uint32_t id)
|
||||
|
||||
string CompilerGLSL::to_qualifiers_glsl(uint32_t id)
|
||||
{
|
||||
auto flags = ir.meta[id].decoration.decoration_flags;
|
||||
auto &flags = ir.meta[id].decoration.decoration_flags;
|
||||
string res;
|
||||
|
||||
auto *var = maybe_get<SPIRVariable>(id);
|
||||
@ -8870,7 +8877,7 @@ string CompilerGLSL::variable_decl(const SPIRVariable &variable)
|
||||
|
||||
const char *CompilerGLSL::to_pls_qualifiers_glsl(const SPIRVariable &variable)
|
||||
{
|
||||
auto flags = ir.meta[variable.self].decoration.decoration_flags;
|
||||
auto &flags = ir.meta[variable.self].decoration.decoration_flags;
|
||||
if (flags.get(DecorationRelaxedPrecision))
|
||||
return "mediump ";
|
||||
else
|
||||
@ -9301,7 +9308,7 @@ void CompilerGLSL::flatten_buffer_block(uint32_t id)
|
||||
auto &var = get<SPIRVariable>(id);
|
||||
auto &type = get<SPIRType>(var.basetype);
|
||||
auto name = to_name(type.self, false);
|
||||
auto flags = ir.meta.at(type.self).decoration.decoration_flags;
|
||||
auto &flags = ir.meta.at(type.self).decoration.decoration_flags;
|
||||
|
||||
if (!type.array.empty())
|
||||
SPIRV_CROSS_THROW(name + " is an array of UBOs.");
|
||||
@ -9598,8 +9605,12 @@ void CompilerGLSL::flush_phi(uint32_t from, uint32_t to)
|
||||
{
|
||||
auto &child = get<SPIRBlock>(to);
|
||||
|
||||
for (auto &phi : child.phi_variables)
|
||||
unordered_set<uint32_t> temporary_phi_variables;
|
||||
|
||||
for (auto itr = begin(child.phi_variables); itr != end(child.phi_variables); ++itr)
|
||||
{
|
||||
auto &phi = *itr;
|
||||
|
||||
if (phi.parent == from)
|
||||
{
|
||||
auto &var = get<SPIRVariable>(phi.function_variable);
|
||||
@ -9611,10 +9622,38 @@ void CompilerGLSL::flush_phi(uint32_t from, uint32_t to)
|
||||
{
|
||||
flush_variable_declaration(phi.function_variable);
|
||||
|
||||
// Check if we are going to write to a Phi variable that another statement will read from
|
||||
// as part of another Phi node in our target block.
|
||||
// For this case, we will need to copy phi.function_variable to a temporary, and use that for future reads.
|
||||
// This is judged to be extremely rare, so deal with it here using a simple, but suboptimal algorithm.
|
||||
bool need_saved_temporary =
|
||||
find_if(itr + 1, end(child.phi_variables), [&](const SPIRBlock::Phi &future_phi) -> bool {
|
||||
return future_phi.local_variable == phi.function_variable && future_phi.parent == from;
|
||||
}) != end(child.phi_variables);
|
||||
|
||||
if (need_saved_temporary)
|
||||
{
|
||||
// Need to make sure we declare the phi variable with a copy at the right scope.
|
||||
// We cannot safely declare a temporary here since we might be inside a continue block.
|
||||
if (!var.allocate_temporary_copy)
|
||||
{
|
||||
var.allocate_temporary_copy = true;
|
||||
force_recompile = true;
|
||||
}
|
||||
statement("_", phi.function_variable, "_copy", " = ", to_name(phi.function_variable), ";");
|
||||
temporary_phi_variables.insert(phi.function_variable);
|
||||
}
|
||||
|
||||
// This might be called in continue block, so make sure we
|
||||
// use this to emit ESSL 1.0 compliant increments/decrements.
|
||||
auto lhs = to_expression(phi.function_variable);
|
||||
auto rhs = to_expression(phi.local_variable);
|
||||
|
||||
string rhs;
|
||||
if (temporary_phi_variables.count(phi.local_variable))
|
||||
rhs = join("_", phi.local_variable, "_copy");
|
||||
else
|
||||
rhs = to_expression(phi.local_variable);
|
||||
|
||||
if (!optimize_read_modify_write(get<SPIRType>(var.basetype), lhs, rhs))
|
||||
statement(lhs, " = ", rhs, ";");
|
||||
}
|
||||
@ -10109,14 +10148,8 @@ void CompilerGLSL::flush_undeclared_variables(SPIRBlock &block)
|
||||
{
|
||||
// Enforce declaration order for regression testing purposes.
|
||||
sort(begin(block.dominated_variables), end(block.dominated_variables));
|
||||
|
||||
for (auto &v : block.dominated_variables)
|
||||
{
|
||||
auto &var = get<SPIRVariable>(v);
|
||||
if (var.deferred_declaration)
|
||||
statement(variable_decl(var), ";");
|
||||
var.deferred_declaration = false;
|
||||
}
|
||||
flush_variable_declaration(v);
|
||||
}
|
||||
|
||||
void CompilerGLSL::emit_hoisted_temporaries(vector<pair<uint32_t, uint32_t>> &temporaries)
|
||||
@ -10129,7 +10162,7 @@ void CompilerGLSL::emit_hoisted_temporaries(vector<pair<uint32_t, uint32_t>> &te
|
||||
for (auto &tmp : temporaries)
|
||||
{
|
||||
add_local_variable_name(tmp.second);
|
||||
auto flags = ir.meta[tmp.second].decoration.decoration_flags;
|
||||
auto &flags = ir.meta[tmp.second].decoration.decoration_flags;
|
||||
auto &type = get<SPIRType>(tmp.first);
|
||||
statement(flags_to_precision_qualifiers_glsl(type, flags), variable_decl(type, to_name(tmp.second)), ";");
|
||||
|
||||
|
||||
Loading…
Reference in New Issue
Block a user