SPIRV-Cross-Vulnerable/spirv_cross.hpp

1183 lines
50 KiB
C++
Raw Normal View History

2016-03-02 17:09:16 +00:00
/*
* Copyright 2015-2021 Arm Limited
* SPDX-License-Identifier: Apache-2.0 OR MIT
2016-03-02 17:09:16 +00:00
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
/*
* At your option, you may choose to accept this material under either:
* 1. The Apache License, Version 2.0, found at <http://www.apache.org/licenses/LICENSE-2.0>, or
* 2. The MIT License, found at <http://opensource.org/licenses/MIT>.
*/
#ifndef SPIRV_CROSS_HPP
#define SPIRV_CROSS_HPP
2016-03-02 17:09:16 +00:00
#ifndef SPV_ENABLE_UTILITY_CODE
#define SPV_ENABLE_UTILITY_CODE
#endif
2016-03-02 17:09:16 +00:00
#include "spirv.hpp"
#include "spirv_cfg.hpp"
#include "spirv_cross_parsed_ir.hpp"
2016-03-02 17:09:16 +00:00
namespace SPIRV_CROSS_NAMESPACE
2016-03-02 17:09:16 +00:00
{
struct Resource
{
// Resources are identified with their SPIR-V ID.
// This is the ID of the OpVariable.
ID id;
// The type ID of the variable which includes arrays and all type modifications.
// This type ID is not suitable for parsing OpMemberDecoration of a struct and other decorations in general
// since these modifications typically happen on the base_type_id.
TypeID type_id;
// The base type of the declared resource.
// This type is the base type which ignores pointers and arrays of the type_id.
// This is mostly useful to parse decorations of the underlying type.
// base_type_id can also be obtained with get_type(get_type(type_id).self).
TypeID base_type_id;
// The declared name (OpName) of the resource.
// For Buffer blocks, the name actually reflects the externally
// visible Block name.
//
// This name can be retrieved again by using either
// get_name(id) or get_name(base_type_id) depending if it's a buffer block or not.
//
// This name can be an empty string in which case get_fallback_name(id) can be
// used which obtains a suitable fallback identifier for an ID.
std::string name;
};
struct BuiltInResource
{
// This is mostly here to support reflection of builtins such as Position/PointSize/CullDistance/ClipDistance.
// This needs to be different from Resource since we can collect builtins from blocks.
// A builtin present here does not necessarily mean it's considered an active builtin,
// since variable ID "activeness" is only tracked on OpVariable level, not Block members.
// For that, update_active_builtins() -> has_active_builtin() can be used to further refine the reflection.
spv::BuiltIn builtin;
// This is the actual value type of the builtin.
// Typically float4, float, array<float, N> for the gl_PerVertex builtins.
// If the builtin is a control point, the control point array type will be stripped away here as appropriate.
TypeID value_type_id;
// This refers to the base resource which contains the builtin.
// If resource is a Block, it can hold multiple builtins, or it might not be a block.
// For advanced reflection scenarios, all information in builtin/value_type_id can be deduced,
// it's just more convenient this way.
Resource resource;
};
struct ShaderResources
{
SmallVector<Resource> uniform_buffers;
SmallVector<Resource> storage_buffers;
SmallVector<Resource> stage_inputs;
SmallVector<Resource> stage_outputs;
SmallVector<Resource> subpass_inputs;
SmallVector<Resource> storage_images;
SmallVector<Resource> sampled_images;
SmallVector<Resource> atomic_counters;
SmallVector<Resource> acceleration_structures;
SmallVector<Resource> gl_plain_uniforms;
// There can only be one push constant block,
// but keep the vector in case this restriction is lifted in the future.
SmallVector<Resource> push_constant_buffers;
SmallVector<Resource> shader_record_buffers;
// For Vulkan GLSL and HLSL source,
// these correspond to separate texture2D and samplers respectively.
SmallVector<Resource> separate_images;
SmallVector<Resource> separate_samplers;
SmallVector<BuiltInResource> builtin_inputs;
SmallVector<BuiltInResource> builtin_outputs;
};
struct CombinedImageSampler
{
// The ID of the sampler2D variable.
VariableID combined_id;
// The ID of the texture2D variable.
VariableID image_id;
// The ID of the sampler variable.
VariableID sampler_id;
};
struct SpecializationConstant
{
// The ID of the specialization constant.
ConstantID id;
// The constant ID of the constant, used in Vulkan during pipeline creation.
uint32_t constant_id;
};
struct BufferRange
{
unsigned index;
size_t offset;
size_t range;
};
enum BufferPackingStandard
{
BufferPackingStd140,
BufferPackingStd430,
BufferPackingStd140EnhancedLayout,
BufferPackingStd430EnhancedLayout,
BufferPackingHLSLCbuffer,
BufferPackingHLSLCbufferPackOffset,
BufferPackingScalar,
BufferPackingScalarEnhancedLayout
};
struct EntryPoint
{
std::string name;
spv::ExecutionModel execution_model;
};
class Compiler
{
public:
friend class CFG;
friend class DominatorBuilder;
// The constructor takes a buffer of SPIR-V words and parses it.
// It will create its own parser, parse the SPIR-V and move the parsed IR
// as if you had called the constructors taking ParsedIR directly.
explicit Compiler(std::vector<uint32_t> ir);
Compiler(const uint32_t *ir, size_t word_count);
// This is more modular. We can also consume a ParsedIR structure directly, either as a move, or copy.
// With copy, we can reuse the same parsed IR for multiple Compiler instances.
explicit Compiler(const ParsedIR &ir);
explicit Compiler(ParsedIR &&ir);
virtual ~Compiler() = default;
// After parsing, API users can modify the SPIR-V via reflection and call this
// to disassemble the SPIR-V into the desired langauage.
// Sub-classes actually implement this.
virtual std::string compile();
// Gets the identifier (OpName) of an ID. If not defined, an empty string will be returned.
const std::string &get_name(ID id) const;
// Applies a decoration to an ID. Effectively injects OpDecorate.
void set_decoration(ID id, spv::Decoration decoration, uint32_t argument = 0);
void set_decoration_string(ID id, spv::Decoration decoration, const std::string &argument);
// Overrides the identifier OpName of an ID.
// Identifiers beginning with underscores or identifiers which contain double underscores
// are reserved by the implementation.
void set_name(ID id, const std::string &name);
// Gets a bitmask for the decorations which are applied to ID.
// I.e. (1ull << spv::DecorationFoo) | (1ull << spv::DecorationBar)
const Bitset &get_decoration_bitset(ID id) const;
// Returns whether the decoration has been applied to the ID.
bool has_decoration(ID id, spv::Decoration decoration) const;
// Gets the value for decorations which take arguments.
// If the decoration is a boolean (i.e. spv::DecorationNonWritable),
// 1 will be returned.
// If decoration doesn't exist or decoration is not recognized,
// 0 will be returned.
uint32_t get_decoration(ID id, spv::Decoration decoration) const;
const std::string &get_decoration_string(ID id, spv::Decoration decoration) const;
// Removes the decoration for an ID.
void unset_decoration(ID id, spv::Decoration decoration);
// Gets the SPIR-V type associated with ID.
// Mostly used with Resource::type_id and Resource::base_type_id to parse the underlying type of a resource.
const SPIRType &get_type(TypeID id) const;
// Gets the SPIR-V type of a variable.
const SPIRType &get_type_from_variable(VariableID id) const;
2016-09-24 20:17:01 +00:00
// Gets the underlying storage class for an OpVariable.
spv::StorageClass get_storage_class(VariableID id) const;
// If get_name() is an empty string, get the fallback name which will be used
// instead in the disassembled source.
virtual const std::string get_fallback_name(ID id) const;
// If get_name() of a Block struct is an empty string, get the fallback name.
// This needs to be per-variable as multiple variables can use the same block type.
virtual const std::string get_block_fallback_name(VariableID id) const;
// Given an OpTypeStruct in ID, obtain the identifier for member number "index".
// This may be an empty string.
const std::string &get_member_name(TypeID id, uint32_t index) const;
// Given an OpTypeStruct in ID, obtain the OpMemberDecoration for member number "index".
uint32_t get_member_decoration(TypeID id, uint32_t index, spv::Decoration decoration) const;
const std::string &get_member_decoration_string(TypeID id, uint32_t index, spv::Decoration decoration) const;
// Sets the member identifier for OpTypeStruct ID, member number "index".
void set_member_name(TypeID id, uint32_t index, const std::string &name);
// Returns the qualified member identifier for OpTypeStruct ID, member number "index",
// or an empty string if no qualified alias exists
const std::string &get_member_qualified_name(TypeID type_id, uint32_t index) const;
// Gets the decoration mask for a member of a struct, similar to get_decoration_mask.
const Bitset &get_member_decoration_bitset(TypeID id, uint32_t index) const;
// Returns whether the decoration has been applied to a member of a struct.
bool has_member_decoration(TypeID id, uint32_t index, spv::Decoration decoration) const;
// Similar to set_decoration, but for struct members.
void set_member_decoration(TypeID id, uint32_t index, spv::Decoration decoration, uint32_t argument = 0);
void set_member_decoration_string(TypeID id, uint32_t index, spv::Decoration decoration,
const std::string &argument);
// Unsets a member decoration, similar to unset_decoration.
void unset_member_decoration(TypeID id, uint32_t index, spv::Decoration decoration);
// Gets the fallback name for a member, similar to get_fallback_name.
virtual const std::string get_fallback_member_name(uint32_t index) const
{
return join("_", index);
}
// Returns a vector of which members of a struct are potentially in use by a
// SPIR-V shader. The granularity of this analysis is per-member of a struct.
// This can be used for Buffer (UBO), BufferBlock/StorageBuffer (SSBO) and PushConstant blocks.
// ID is the Resource::id obtained from get_shader_resources().
SmallVector<BufferRange> get_active_buffer_ranges(VariableID id) const;
// Returns the effective size of a buffer block.
size_t get_declared_struct_size(const SPIRType &struct_type) const;
// Returns the effective size of a buffer block, with a given array size
// for a runtime array.
// SSBOs are typically declared as runtime arrays. get_declared_struct_size() will return 0 for the size.
// This is not very helpful for applications which might need to know the array stride of its last member.
// This can be done through the API, but it is not very intuitive how to accomplish this, so here we provide a helper function
// to query the size of the buffer, assuming that the last member has a certain size.
// If the buffer does not contain a runtime array, array_size is ignored, and the function will behave as
// get_declared_struct_size().
// To get the array stride of the last member, something like:
// get_declared_struct_size_runtime_array(type, 1) - get_declared_struct_size_runtime_array(type, 0) will work.
size_t get_declared_struct_size_runtime_array(const SPIRType &struct_type, size_t array_size) const;
// Returns the effective size of a buffer block struct member.
size_t get_declared_struct_member_size(const SPIRType &struct_type, uint32_t index) const;
2016-08-26 10:58:50 +00:00
// Returns a set of all global variables which are statically accessed
// by the control flow graph from the current entry point.
// Only variables which change the interface for a shader are returned, that is,
// variables with storage class of Input, Output, Uniform, UniformConstant, PushConstant and AtomicCounter
// storage classes are returned.
//
// To use the returned set as the filter for which variables are used during compilation,
// this set can be moved to set_enabled_interface_variables().
std::unordered_set<VariableID> get_active_interface_variables() const;
2016-08-26 10:58:50 +00:00
// Sets the interface variables which are used during compilation.
// By default, all variables are used.
// Once set, compile() will only consider the set in active_variables.
void set_enabled_interface_variables(std::unordered_set<VariableID> active_variables);
2016-08-26 10:58:50 +00:00
// Query shader resources, use ids with reflection interface to modify or query binding points, etc.
ShaderResources get_shader_resources() const;
2016-08-26 10:58:50 +00:00
// Query shader resources, but only return the variables which are part of active_variables.
// E.g.: get_shader_resources(get_active_variables()) to only return the variables which are statically
// accessed.
ShaderResources get_shader_resources(const std::unordered_set<VariableID> &active_variables) const;
2016-08-26 10:58:50 +00:00
// Remapped variables are considered built-in variables and a backend will
// not emit a declaration for this variable.
// This is mostly useful for making use of builtins which are dependent on extensions.
void set_remapped_variable_state(VariableID id, bool remap_enable);
bool get_remapped_variable_state(VariableID id) const;
// For subpassInput variables which are remapped to plain variables,
// the number of components in the remapped
// variable must be specified as the backing type of subpass inputs are opaque.
void set_subpass_input_remapped_components(VariableID id, uint32_t components);
uint32_t get_subpass_input_remapped_components(VariableID id) const;
// All operations work on the current entry point.
// Entry points can be swapped out with set_entry_point().
// Entry points should be set right after the constructor completes as some reflection functions traverse the graph from the entry point.
// Resource reflection also depends on the entry point.
// By default, the current entry point is set to the first OpEntryPoint which appears in the SPIR-V module.
2017-11-10 21:40:33 +00:00
// Some shader languages restrict the names that can be given to entry points, and the
// corresponding backend will automatically rename an entry point name, during the call
// to compile() if it is illegal. For example, the common entry point name main() is
// illegal in MSL, and is renamed to an alternate name by the MSL backend.
// Given the original entry point name contained in the SPIR-V, this function returns
// the name, as updated by the backend during the call to compile(). If the name is not
// illegal, and has not been renamed, or if this function is called before compile(),
// this function will simply return the same name.
// New variants of entry point query and reflection.
// Names for entry points in the SPIR-V module may alias if they belong to different execution models.
// To disambiguate, we must pass along with the entry point names the execution model.
SmallVector<EntryPoint> get_entry_points_and_stages() const;
void set_entry_point(const std::string &entry, spv::ExecutionModel execution_model);
// Renames an entry point from old_name to new_name.
// If old_name is currently selected as the current entry point, it will continue to be the current entry point,
// albeit with a new name.
// get_entry_points() is essentially invalidated at this point.
2018-03-01 13:03:59 +00:00
void rename_entry_point(const std::string &old_name, const std::string &new_name,
spv::ExecutionModel execution_model);
const SPIREntryPoint &get_entry_point(const std::string &name, spv::ExecutionModel execution_model) const;
SPIREntryPoint &get_entry_point(const std::string &name, spv::ExecutionModel execution_model);
2018-03-01 13:03:59 +00:00
const std::string &get_cleansed_entry_point_name(const std::string &name,
spv::ExecutionModel execution_model) const;
// Traverses all reachable opcodes and sets active_builtins to a bitmask of all builtin variables which are accessed in the shader.
void update_active_builtins();
bool has_active_builtin(spv::BuiltIn builtin, spv::StorageClass storage) const;
// Query and modify OpExecutionMode.
const Bitset &get_execution_mode_bitset() const;
void unset_execution_mode(spv::ExecutionMode mode);
void set_execution_mode(spv::ExecutionMode mode, uint32_t arg0 = 0, uint32_t arg1 = 0, uint32_t arg2 = 0);
// Gets argument for an execution mode (LocalSize, Invocations, OutputVertices).
// For LocalSize or LocalSizeId, the index argument is used to select the dimension (X = 0, Y = 1, Z = 2).
// For execution modes which do not have arguments, 0 is returned.
// LocalSizeId query returns an ID. If LocalSizeId execution mode is not used, it returns 0.
// LocalSize always returns a literal. If execution mode is LocalSizeId,
// the literal (spec constant or not) is still returned.
uint32_t get_execution_mode_argument(spv::ExecutionMode mode, uint32_t index = 0) const;
spv::ExecutionModel get_execution_model() const;
bool is_tessellation_shader() const;
bool is_tessellating_triangles() const;
// In SPIR-V, the compute work group size can be represented by a constant vector, in which case
// the LocalSize execution mode is ignored.
//
// This constant vector can be a constant vector, specialization constant vector, or partly specialized constant vector.
// To modify and query work group dimensions which are specialization constants, SPIRConstant values must be modified
// directly via get_constant() rather than using LocalSize directly. This function will return which constants should be modified.
//
// To modify dimensions which are *not* specialization constants, set_execution_mode should be used directly.
// Arguments to set_execution_mode which are specialization constants are effectively ignored during compilation.
// NOTE: This is somewhat different from how SPIR-V works. In SPIR-V, the constant vector will completely replace LocalSize,
// while in this interface, LocalSize is only ignored for specialization constants.
//
// The specialization constant will be written to x, y and z arguments.
// If the component is not a specialization constant, a zeroed out struct will be written.
// The return value is the constant ID of the builtin WorkGroupSize, but this is not expected to be useful
// for most use cases.
// If LocalSizeId is used, there is no uvec3 value representing the workgroup size, so the return value is 0,
// but x, y and z are written as normal if the components are specialization constants.
uint32_t get_work_group_size_specialization_constants(SpecializationConstant &x, SpecializationConstant &y,
SpecializationConstant &z) const;
// Analyzes all OpImageFetch (texelFetch) opcodes and checks if there are instances where
// said instruction is used without a combined image sampler.
// GLSL targets do not support the use of texelFetch without a sampler.
// To workaround this, we must inject a dummy sampler which can be used to form a sampler2D at the call-site of
// texelFetch as necessary.
//
// This must be called before build_combined_image_samplers().
// build_combined_image_samplers() may refer to the ID returned by this method if the returned ID is non-zero.
// The return value will be the ID of a sampler object if a dummy sampler is necessary, or 0 if no sampler object
// is required.
//
// If the returned ID is non-zero, it can be decorated with set/bindings as desired before calling compile().
// Calling this function also invalidates get_active_interface_variables(), so this should be called
// before that function.
VariableID build_dummy_sampler_for_combined_images();
// Analyzes all separate image and samplers used from the currently selected entry point,
// and re-routes them all to a combined image sampler instead.
// This is required to "support" separate image samplers in targets which do not natively support
// this feature, like GLSL/ESSL.
//
// This must be called before compile() if such remapping is desired.
// This call will add new sampled images to the SPIR-V,
// so it will appear in reflection if get_shader_resources() is called after build_combined_image_samplers.
//
// If any image/sampler remapping was found, no separate image/samplers will appear in the decompiled output,
// but will still appear in reflection.
//
// The resulting samplers will be void of any decorations like name, descriptor sets and binding points,
// so this can be added before compile() if desired.
//
// Combined image samplers originating from this set are always considered active variables.
// Arrays of separate samplers are not supported, but arrays of separate images are supported.
// Array of images + sampler -> Array of combined image samplers.
void build_combined_image_samplers();
// Gets a remapping for the combined image samplers.
const SmallVector<CombinedImageSampler> &get_combined_image_samplers() const
{
return combined_image_samplers;
}
2016-09-20 08:17:41 +00:00
// Set a new variable type remap callback.
// The type remapping is designed to allow global interface variable to assume more special types.
// A typical example here is to remap sampler2D into samplerExternalOES, which currently isn't supported
// directly by SPIR-V.
//
// In compile() while emitting code,
// for every variable that is declared, including function parameters, the callback will be called
// and the API user has a chance to change the textual representation of the type used to declare the variable.
// The API user can detect special patterns in names to guide the remapping.
2016-09-20 08:17:41 +00:00
void set_variable_type_remap_callback(VariableTypeRemapCallback cb)
{
variable_remap_callback = std::move(cb);
}
// API for querying which specialization constants exist.
// To modify a specialization constant before compile(), use get_constant(constant.id),
// then update constants directly in the SPIRConstant data structure.
// For composite types, the subconstants can be iterated over and modified.
// constant_type is the SPIRType for the specialization constant,
// which can be queried to determine which fields in the unions should be poked at.
SmallVector<SpecializationConstant> get_specialization_constants() const;
SPIRConstant &get_constant(ConstantID id);
const SPIRConstant &get_constant(ConstantID id) const;
uint32_t get_current_id_bound() const
{
return uint32_t(ir.ids.size());
}
// API for querying buffer objects.
// The type passed in here should be the base type of a resource, i.e.
// get_type(resource.base_type_id)
// as decorations are set in the basic Block type.
// The type passed in here must have these decorations set, or an exception is raised.
// Only UBOs and SSBOs or sub-structs which are part of these buffer types will have these decorations set.
uint32_t type_struct_member_offset(const SPIRType &type, uint32_t index) const;
uint32_t type_struct_member_array_stride(const SPIRType &type, uint32_t index) const;
uint32_t type_struct_member_matrix_stride(const SPIRType &type, uint32_t index) const;
// Gets the offset in SPIR-V words (uint32_t) for a decoration which was originally declared in the SPIR-V binary.
// The offset will point to one or more uint32_t literals which can be modified in-place before using the SPIR-V binary.
// Note that adding or removing decorations using the reflection API will not change the behavior of this function.
// If the decoration was declared, sets the word_offset to an offset into the provided SPIR-V binary buffer and returns true,
// otherwise, returns false.
// If the decoration does not have any value attached to it (e.g. DecorationRelaxedPrecision), this function will also return false.
bool get_binary_offset_for_decoration(VariableID id, spv::Decoration decoration, uint32_t &word_offset) const;
// HLSL counter buffer reflection interface.
// Append/Consume/Increment/Decrement in HLSL is implemented as two "neighbor" buffer objects where
// one buffer implements the storage, and a single buffer containing just a lone "int" implements the counter.
// To SPIR-V these will be exposed as two separate buffers, but glslang HLSL frontend emits a special indentifier
// which lets us link the two buffers together.
// Queries if a variable ID is a counter buffer which "belongs" to a regular buffer object.
// If SPV_GOOGLE_hlsl_functionality1 is used, this can be used even with a stripped SPIR-V module.
// Otherwise, this query is purely based on OpName identifiers as found in the SPIR-V module, and will
// only return true if OpSource was reported HLSL.
// To rely on this functionality, ensure that the SPIR-V module is not stripped.
bool buffer_is_hlsl_counter_buffer(VariableID id) const;
// Queries if a buffer object has a neighbor "counter" buffer.
// If so, the ID of that counter buffer will be returned in counter_id.
// If SPV_GOOGLE_hlsl_functionality1 is used, this can be used even with a stripped SPIR-V module.
// Otherwise, this query is purely based on OpName identifiers as found in the SPIR-V module, and will
// only return true if OpSource was reported HLSL.
// To rely on this functionality, ensure that the SPIR-V module is not stripped.
bool buffer_get_hlsl_counter_buffer(VariableID id, uint32_t &counter_id) const;
// Gets the list of all SPIR-V Capabilities which were declared in the SPIR-V module.
const SmallVector<spv::Capability> &get_declared_capabilities() const;
// Gets the list of all SPIR-V extensions which were declared in the SPIR-V module.
const SmallVector<std::string> &get_declared_extensions() const;
// When declaring buffer blocks in GLSL, the name declared in the GLSL source
// might not be the same as the name declared in the SPIR-V module due to naming conflicts.
// In this case, SPIRV-Cross needs to find a fallback-name, and it might only
// be possible to know this name after compiling to GLSL.
// This is particularly important for HLSL input and UAVs which tends to reuse the same block type
// for multiple distinct blocks. For these cases it is not possible to modify the name of the type itself
// because it might be unique. Instead, you can use this interface to check after compilation which
// name was actually used if your input SPIR-V tends to have this problem.
// For other names like remapped names for variables, etc, it's generally enough to query the name of the variables
// after compiling, block names are an exception to this rule.
// ID is the name of a variable as returned by Resource::id, and must be a variable with a Block-like type.
//
// This also applies to HLSL cbuffers.
std::string get_remapped_declared_block_name(VariableID id) const;
// For buffer block variables, get the decorations for that variable.
// Sometimes, decorations for buffer blocks are found in member decorations instead
// of direct decorations on the variable itself.
// The most common use here is to check if a buffer is readonly or writeonly.
Bitset get_buffer_block_flags(VariableID id) const;
2019-09-23 22:05:04 +00:00
// Returns whether the position output is invariant
bool is_position_invariant() const
{
return position_invariant;
}
protected:
const uint32_t *stream(const Instruction &instr) const
{
// If we're not going to use any arguments, just return nullptr.
// We want to avoid case where we return an out of range pointer
// that trips debug assertions on some platforms.
if (!instr.length)
return nullptr;
if (instr.is_embedded())
{
auto &embedded = static_cast<const EmbeddedInstruction &>(instr);
assert(embedded.ops.size() == instr.length);
return embedded.ops.data();
}
else
{
if (instr.offset + instr.length > ir.spirv.size())
SPIRV_CROSS_THROW("Compiler::stream() out of range.");
return &ir.spirv[instr.offset];
}
}
uint32_t *stream_mutable(const Instruction &instr) const
{
return const_cast<uint32_t *>(stream(instr));
}
ParsedIR ir;
// Marks variables which have global scope and variables which can alias with other variables
// (SSBO, image load store, etc)
SmallVector<uint32_t> global_variables;
SmallVector<uint32_t> aliased_variables;
SPIRFunction *current_function = nullptr;
SPIRBlock *current_block = nullptr;
uint32_t current_loop_level = 0;
std::unordered_set<VariableID> active_interface_variables;
2016-08-26 10:58:50 +00:00
bool check_active_interface_variables = false;
void add_loop_level();
void set_initializers(SPIRExpression &e)
{
e.emitted_loop_level = current_loop_level;
}
template <typename T>
void set_initializers(const T &)
{
}
// If our IDs are out of range here as part of opcodes, throw instead of
// undefined behavior.
template <typename T, typename... P>
T &set(uint32_t id, P &&... args)
{
ir.add_typed_id(static_cast<Types>(T::type), id);
auto &var = variant_set<T>(ir.ids[id], std::forward<P>(args)...);
var.self = id;
set_initializers(var);
return var;
}
template <typename T>
T &get(uint32_t id)
{
return variant_get<T>(ir.ids[id]);
}
template <typename T>
T *maybe_get(uint32_t id)
{
if (id >= ir.ids.size())
return nullptr;
else if (ir.ids[id].get_type() == static_cast<Types>(T::type))
return &get<T>(id);
else
return nullptr;
}
template <typename T>
const T &get(uint32_t id) const
{
return variant_get<T>(ir.ids[id]);
}
template <typename T>
const T *maybe_get(uint32_t id) const
{
if (id >= ir.ids.size())
return nullptr;
else if (ir.ids[id].get_type() == static_cast<Types>(T::type))
return &get<T>(id);
else
return nullptr;
}
// Gets the id of SPIR-V type underlying the given type_id, which might be a pointer.
uint32_t get_pointee_type_id(uint32_t type_id) const;
// Gets the SPIR-V type underlying the given type, which might be a pointer.
const SPIRType &get_pointee_type(const SPIRType &type) const;
// Gets the SPIR-V type underlying the given type_id, which might be a pointer.
const SPIRType &get_pointee_type(uint32_t type_id) const;
// Gets the ID of the SPIR-V type underlying a variable.
uint32_t get_variable_data_type_id(const SPIRVariable &var) const;
// Gets the SPIR-V type underlying a variable.
SPIRType &get_variable_data_type(const SPIRVariable &var);
// Gets the SPIR-V type underlying a variable.
const SPIRType &get_variable_data_type(const SPIRVariable &var) const;
// Gets the SPIR-V element type underlying an array variable.
SPIRType &get_variable_element_type(const SPIRVariable &var);
// Gets the SPIR-V element type underlying an array variable.
const SPIRType &get_variable_element_type(const SPIRVariable &var) const;
// Sets the qualified member identifier for OpTypeStruct ID, member number "index".
void set_member_qualified_name(uint32_t type_id, uint32_t index, const std::string &name);
void set_qualified_name(uint32_t id, const std::string &name);
// Returns if the given type refers to a sampled image.
bool is_sampled_image_type(const SPIRType &type);
const SPIREntryPoint &get_entry_point() const;
SPIREntryPoint &get_entry_point();
static bool is_tessellation_shader(spv::ExecutionModel model);
virtual std::string to_name(uint32_t id, bool allow_alias = true) const;
bool is_builtin_variable(const SPIRVariable &var) const;
bool is_builtin_type(const SPIRType &type) const;
2016-08-26 10:58:50 +00:00
bool is_hidden_variable(const SPIRVariable &var, bool include_builtins = false) const;
bool is_immutable(uint32_t id) const;
bool is_member_builtin(const SPIRType &type, uint32_t index, spv::BuiltIn *builtin) const;
bool is_scalar(const SPIRType &type) const;
bool is_vector(const SPIRType &type) const;
bool is_matrix(const SPIRType &type) const;
bool is_array(const SPIRType &type) const;
bool is_pointer(const SPIRType &type) const;
bool is_physical_pointer(const SPIRType &type) const;
bool is_physical_pointer_to_buffer_block(const SPIRType &type) const;
static bool is_runtime_size_array(const SPIRType &type);
uint32_t expression_type_id(uint32_t id) const;
const SPIRType &expression_type(uint32_t id) const;
bool expression_is_lvalue(uint32_t id) const;
bool variable_storage_is_aliased(const SPIRVariable &var);
SPIRVariable *maybe_get_backing_variable(uint32_t chain);
void register_read(uint32_t expr, uint32_t chain, bool forwarded);
void register_write(uint32_t chain);
inline bool is_continue(uint32_t next) const
{
return (ir.block_meta[next] & ParsedIR::BLOCK_META_CONTINUE_BIT) != 0;
}
inline bool is_single_block_loop(uint32_t next) const
{
auto &block = get<SPIRBlock>(next);
return block.merge == SPIRBlock::MergeLoop && block.continue_block == ID(next);
}
inline bool is_break(uint32_t next) const
{
return (ir.block_meta[next] &
(ParsedIR::BLOCK_META_LOOP_MERGE_BIT | ParsedIR::BLOCK_META_MULTISELECT_MERGE_BIT)) != 0;
}
inline bool is_loop_break(uint32_t next) const
{
return (ir.block_meta[next] & ParsedIR::BLOCK_META_LOOP_MERGE_BIT) != 0;
}
inline bool is_conditional(uint32_t next) const
{
return (ir.block_meta[next] &
(ParsedIR::BLOCK_META_SELECTION_MERGE_BIT | ParsedIR::BLOCK_META_MULTISELECT_MERGE_BIT)) != 0;
}
// Dependency tracking for temporaries read from variables.
void flush_dependees(SPIRVariable &var);
void flush_all_active_variables();
void flush_control_dependent_expressions(uint32_t block);
void flush_all_atomic_capable_variables();
void flush_all_aliased_variables();
void register_global_read_dependencies(const SPIRBlock &func, uint32_t id);
void register_global_read_dependencies(const SPIRFunction &func, uint32_t id);
std::unordered_set<uint32_t> invalid_expressions;
void update_name_cache(std::unordered_set<std::string> &cache, std::string &name);
// A variant which takes two sets of names. The secondary is only used to verify there are no collisions,
// but the set is not updated when we have found a new name.
// Used primarily when adding block interface names.
void update_name_cache(std::unordered_set<std::string> &cache_primary,
2019-01-07 09:01:28 +00:00
const std::unordered_set<std::string> &cache_secondary, std::string &name);
bool function_is_pure(const SPIRFunction &func);
bool block_is_pure(const SPIRBlock &block);
bool function_is_control_dependent(const SPIRFunction &func);
bool block_is_control_dependent(const SPIRBlock &block);
bool execution_is_branchless(const SPIRBlock &from, const SPIRBlock &to) const;
bool execution_is_direct_branch(const SPIRBlock &from, const SPIRBlock &to) const;
bool execution_is_noop(const SPIRBlock &from, const SPIRBlock &to) const;
SPIRBlock::ContinueBlockType continue_block_type(const SPIRBlock &continue_block) const;
void force_recompile();
void force_recompile_guarantee_forward_progress();
void clear_force_recompile();
bool is_forcing_recompilation() const;
bool is_force_recompile = false;
bool is_force_recompile_forward_progress = false;
2023-01-10 02:14:37 +00:00
bool block_is_noop(const SPIRBlock &block) const;
bool block_is_loop_candidate(const SPIRBlock &block, SPIRBlock::Method method) const;
bool types_are_logically_equivalent(const SPIRType &a, const SPIRType &b) const;
void inherit_expression_dependencies(uint32_t dst, uint32_t source);
void add_implied_read_expression(SPIRExpression &e, uint32_t source);
void add_implied_read_expression(SPIRAccessChain &e, uint32_t source);
void add_active_interface_variable(uint32_t var_id);
// For proper multiple entry point support, allow querying if an Input or Output
// variable is part of that entry points interface.
bool interface_variable_exists_in_entry_point(uint32_t id) const;
SmallVector<CombinedImageSampler> combined_image_samplers;
2019-09-23 22:05:04 +00:00
void remap_variable_type_name(const SPIRType &type, const std::string &var_name, std::string &type_name) const
2016-09-20 08:17:41 +00:00
{
if (variable_remap_callback)
variable_remap_callback(type, var_name, type_name);
}
void set_ir(const ParsedIR &parsed);
void set_ir(ParsedIR &&parsed);
void parse_fixup();
// Used internally to implement various traversals for queries.
struct OpcodeHandler
{
virtual ~OpcodeHandler() = default;
// Return true if traversal should continue.
// If false, traversal will end immediately.
virtual bool handle(spv::Op opcode, const uint32_t *args, uint32_t length) = 0;
virtual bool handle_terminator(const SPIRBlock &)
{
return true;
}
virtual bool follow_function_call(const SPIRFunction &)
{
return true;
}
virtual void set_current_block(const SPIRBlock &)
{
}
// Called after returning from a function or when entering a block,
// can be called multiple times per block,
// while set_current_block is only called on block entry.
virtual void rearm_current_block(const SPIRBlock &)
{
}
virtual bool begin_function_scope(const uint32_t *, uint32_t)
{
return true;
}
virtual bool end_function_scope(const uint32_t *, uint32_t)
{
return true;
}
};
struct BufferAccessHandler : OpcodeHandler
{
BufferAccessHandler(const Compiler &compiler_, SmallVector<BufferRange> &ranges_, uint32_t id_)
: compiler(compiler_)
, ranges(ranges_)
, id(id_)
{
}
bool handle(spv::Op opcode, const uint32_t *args, uint32_t length) override;
const Compiler &compiler;
SmallVector<BufferRange> &ranges;
uint32_t id;
std::unordered_set<uint32_t> seen;
};
2016-08-26 10:58:50 +00:00
struct InterfaceVariableAccessHandler : OpcodeHandler
{
InterfaceVariableAccessHandler(const Compiler &compiler_, std::unordered_set<VariableID> &variables_)
2016-08-26 10:58:50 +00:00
: compiler(compiler_)
, variables(variables_)
{
}
bool handle(spv::Op opcode, const uint32_t *args, uint32_t length) override;
const Compiler &compiler;
std::unordered_set<VariableID> &variables;
2016-08-26 10:58:50 +00:00
};
struct CombinedImageSamplerHandler : OpcodeHandler
{
CombinedImageSamplerHandler(Compiler &compiler_)
: compiler(compiler_)
{
}
bool handle(spv::Op opcode, const uint32_t *args, uint32_t length) override;
bool begin_function_scope(const uint32_t *args, uint32_t length) override;
bool end_function_scope(const uint32_t *args, uint32_t length) override;
Compiler &compiler;
// Each function in the call stack needs its own remapping for parameters so we can deduce which global variable each texture/sampler the parameter is statically bound to.
std::stack<std::unordered_map<uint32_t, uint32_t>> parameter_remapping;
std::stack<SPIRFunction *> functions;
uint32_t remap_parameter(uint32_t id);
void push_remap_parameters(const SPIRFunction &func, const uint32_t *args, uint32_t length);
void pop_remap_parameters();
void register_combined_image_sampler(SPIRFunction &caller, VariableID combined_id, VariableID texture_id,
VariableID sampler_id, bool depth);
};
struct DummySamplerForCombinedImageHandler : OpcodeHandler
{
DummySamplerForCombinedImageHandler(Compiler &compiler_)
: compiler(compiler_)
{
}
bool handle(spv::Op opcode, const uint32_t *args, uint32_t length) override;
Compiler &compiler;
bool need_dummy_sampler = false;
};
struct ActiveBuiltinHandler : OpcodeHandler
{
ActiveBuiltinHandler(Compiler &compiler_)
2017-03-06 15:50:46 +00:00
: compiler(compiler_)
{
}
bool handle(spv::Op opcode, const uint32_t *args, uint32_t length) override;
Compiler &compiler;
void handle_builtin(const SPIRType &type, spv::BuiltIn builtin, const Bitset &decoration_flags);
void add_if_builtin(uint32_t id);
void add_if_builtin_or_block(uint32_t id);
void add_if_builtin(uint32_t id, bool allow_blocks);
};
bool traverse_all_reachable_opcodes(const SPIRBlock &block, OpcodeHandler &handler) const;
bool traverse_all_reachable_opcodes(const SPIRFunction &block, OpcodeHandler &handler) const;
// This must be an ordered data structure so we always pick the same type aliases.
SmallVector<uint32_t> global_struct_cache;
2016-08-26 10:58:50 +00:00
ShaderResources get_shader_resources(const std::unordered_set<VariableID> *active_variables) const;
2016-09-20 08:17:41 +00:00
VariableTypeRemapCallback variable_remap_callback;
2017-01-21 09:07:38 +00:00
bool get_common_basic_type(const SPIRType &type, SPIRType::BaseType &base_type);
std::unordered_set<uint32_t> forced_temporaries;
std::unordered_set<uint32_t> forwarded_temporaries;
std::unordered_set<uint32_t> suppressed_usage_tracking;
std::unordered_set<uint32_t> hoisted_temporaries;
std::unordered_set<uint32_t> forced_invariant_temporaries;
Bitset active_input_builtins;
Bitset active_output_builtins;
uint32_t clip_distance_count = 0;
uint32_t cull_distance_count = 0;
2018-03-01 11:30:55 +00:00
bool position_invariant = false;
2017-03-25 15:28:44 +00:00
void analyze_parameter_preservation(
SPIRFunction &entry, const CFG &cfg,
const std::unordered_map<uint32_t, std::unordered_set<uint32_t>> &variable_to_blocks,
const std::unordered_map<uint32_t, std::unordered_set<uint32_t>> &complete_write_blocks);
// If a variable ID or parameter ID is found in this set, a sampler is actually a shadow/comparison sampler.
// SPIR-V does not support this distinction, so we must keep track of this information outside the type system.
// There might be unrelated IDs found in this set which do not correspond to actual variables.
// This set should only be queried for the existence of samplers which are already known to be variables or parameter IDs.
// Similar is implemented for images, as well as if subpass inputs are needed.
std::unordered_set<uint32_t> comparison_ids;
bool need_subpass_input = false;
bool need_subpass_input_ms = false;
// In certain backends, we will need to use a dummy sampler to be able to emit code.
// GLSL does not support texelFetch on texture2D objects, but SPIR-V does,
// so we need to workaround by having the application inject a dummy sampler.
uint32_t dummy_sampler_id = 0;
void analyze_image_and_sampler_usage();
struct CombinedImageSamplerDrefHandler : OpcodeHandler
{
CombinedImageSamplerDrefHandler(Compiler &compiler_)
: compiler(compiler_)
{
}
bool handle(spv::Op opcode, const uint32_t *args, uint32_t length) override;
Compiler &compiler;
std::unordered_set<uint32_t> dref_combined_samplers;
};
struct CombinedImageSamplerUsageHandler : OpcodeHandler
{
CombinedImageSamplerUsageHandler(Compiler &compiler_,
const std::unordered_set<uint32_t> &dref_combined_samplers_)
: compiler(compiler_)
, dref_combined_samplers(dref_combined_samplers_)
{
}
bool begin_function_scope(const uint32_t *args, uint32_t length) override;
bool handle(spv::Op opcode, const uint32_t *args, uint32_t length) override;
Compiler &compiler;
const std::unordered_set<uint32_t> &dref_combined_samplers;
std::unordered_map<uint32_t, std::unordered_set<uint32_t>> dependency_hierarchy;
std::unordered_set<uint32_t> comparison_ids;
void add_hierarchy_to_comparison_ids(uint32_t ids);
bool need_subpass_input = false;
bool need_subpass_input_ms = false;
void add_dependency(uint32_t dst, uint32_t src);
};
2017-08-03 12:32:07 +00:00
2018-07-05 08:42:05 +00:00
void build_function_control_flow_graphs_and_analyze();
std::unordered_map<uint32_t, std::unique_ptr<CFG>> function_cfgs;
const CFG &get_cfg_for_current_function() const;
const CFG &get_cfg_for_function(uint32_t id) const;
struct CFGBuilder : OpcodeHandler
{
explicit CFGBuilder(Compiler &compiler_);
bool follow_function_call(const SPIRFunction &func) override;
bool handle(spv::Op op, const uint32_t *args, uint32_t length) override;
Compiler &compiler;
std::unordered_map<uint32_t, std::unique_ptr<CFG>> function_cfgs;
};
struct AnalyzeVariableScopeAccessHandler : OpcodeHandler
{
AnalyzeVariableScopeAccessHandler(Compiler &compiler_, SPIRFunction &entry_);
bool follow_function_call(const SPIRFunction &) override;
void set_current_block(const SPIRBlock &block) override;
void notify_variable_access(uint32_t id, uint32_t block);
bool id_is_phi_variable(uint32_t id) const;
bool id_is_potential_temporary(uint32_t id) const;
bool handle(spv::Op op, const uint32_t *args, uint32_t length) override;
bool handle_terminator(const SPIRBlock &block) override;
Compiler &compiler;
SPIRFunction &entry;
std::unordered_map<uint32_t, std::unordered_set<uint32_t>> accessed_variables_to_block;
std::unordered_map<uint32_t, std::unordered_set<uint32_t>> accessed_temporaries_to_block;
std::unordered_map<uint32_t, uint32_t> result_id_to_type;
std::unordered_map<uint32_t, std::unordered_set<uint32_t>> complete_write_variables_to_block;
std::unordered_map<uint32_t, std::unordered_set<uint32_t>> partial_write_variables_to_block;
std::unordered_set<uint32_t> access_chain_expressions;
// Access chains used in multiple blocks mean hoisting all the variables used to construct the access chain as not all backends can use pointers.
// This is also relevant when forwarding opaque objects since we cannot lower these to temporaries.
std::unordered_map<uint32_t, std::unordered_set<uint32_t>> rvalue_forward_children;
const SPIRBlock *current_block = nullptr;
};
struct StaticExpressionAccessHandler : OpcodeHandler
{
StaticExpressionAccessHandler(Compiler &compiler_, uint32_t variable_id_);
bool follow_function_call(const SPIRFunction &) override;
bool handle(spv::Op op, const uint32_t *args, uint32_t length) override;
Compiler &compiler;
uint32_t variable_id;
uint32_t static_expression = 0;
uint32_t write_count = 0;
};
struct PhysicalBlockMeta
{
uint32_t alignment = 0;
};
struct PhysicalStorageBufferPointerHandler : OpcodeHandler
{
explicit PhysicalStorageBufferPointerHandler(Compiler &compiler_);
bool handle(spv::Op op, const uint32_t *args, uint32_t length) override;
Compiler &compiler;
std::unordered_set<uint32_t> non_block_types;
std::unordered_map<uint32_t, PhysicalBlockMeta> physical_block_type_meta;
std::unordered_map<uint32_t, PhysicalBlockMeta *> access_chain_to_physical_block;
void mark_aligned_access(uint32_t id, const uint32_t *args, uint32_t length);
PhysicalBlockMeta *find_block_meta(uint32_t id) const;
bool type_is_bda_block_entry(uint32_t type_id) const;
void setup_meta_chain(uint32_t type_id, uint32_t var_id);
uint32_t get_minimum_scalar_alignment(const SPIRType &type) const;
void analyze_non_block_types_from_block(const SPIRType &type);
uint32_t get_base_non_block_type_id(uint32_t type_id) const;
};
void analyze_non_block_pointer_types();
SmallVector<uint32_t> physical_storage_non_block_pointer_types;
std::unordered_map<uint32_t, PhysicalBlockMeta> physical_storage_type_to_alignment;
void analyze_variable_scope(SPIRFunction &function, AnalyzeVariableScopeAccessHandler &handler);
void find_function_local_luts(SPIRFunction &function, const AnalyzeVariableScopeAccessHandler &handler,
bool single_function);
bool may_read_undefined_variable_in_block(const SPIRBlock &block, uint32_t var);
Support the SPV_EXT_fragment_shader_interlock extension. This was straightforward to implement in GLSL. The `ShadingRateInterlockOrderedEXT` and `ShadingRateInterlockUnorderedEXT` modes aren't implemented yet, because we don't support `SPV_NV_shading_rate` or `SPV_EXT_fragment_invocation_density` yet. HLSL and MSL were more interesting. They don't support this directly, but they do support marking resources as "rasterizer ordered," which does roughly the same thing. So this implementation scans all accesses inside the critical section and marks all storage resources found therein as rasterizer ordered. They also don't support the fine-grained controls on pixel- vs. sample-level interlock and disabling ordering guarantees that GLSL and SPIR-V do, but that's OK. "Unordered" here merely means the order is undefined; that it just so happens to be the same as rasterizer order is immaterial. As for pixel- vs. sample-level interlock, Vulkan explicitly states: > With sample shading enabled, [the `PixelInterlockOrderedEXT` and > `PixelInterlockUnorderedEXT`] execution modes are treated like > `SampleInterlockOrderedEXT` or `SampleInterlockUnorderedEXT` > respectively. and: > If [the `SampleInterlockOrderedEXT` or `SampleInterlockUnorderedEXT`] > execution modes are used in single-sample mode they are treated like > `PixelInterlockOrderedEXT` or `PixelInterlockUnorderedEXT` > respectively. So this will DTRT for MoltenVK and gfx-rs, at least. MSL additionally supports multiple raster order groups; resources that are not accessed together can be placed in different ROGs to allow them to be synchronized separately. A more sophisticated analysis might be able to place resources optimally, but that's outside the scope of this change. For now, we assign all resources to group 0, which should do for our purposes. `glslang` doesn't support the `RasterizerOrdered` UAVs this implementation produces for HLSL, so the test case needs `fxc.exe`. It also insists on GLSL 4.50 for `GL_ARB_fragment_shader_interlock`, even though the spec says it needs either 4.20 or `GL_ARB_shader_image_load_store`; and it doesn't support the `GL_NV_fragment_shader_interlock` extension at all. So I haven't been able to test those code paths. Fixes #1002.
2019-08-04 05:07:20 +00:00
// Finds all resources that are written to from inside the critical section, if present.
// The critical section is delimited by OpBeginInvocationInterlockEXT and
// OpEndInvocationInterlockEXT instructions. In MSL and HLSL, any resources written
// while inside the critical section must be placed in a raster order group.
struct InterlockedResourceAccessHandler : OpcodeHandler
{
InterlockedResourceAccessHandler(Compiler &compiler_, uint32_t entry_point_id)
Support the SPV_EXT_fragment_shader_interlock extension. This was straightforward to implement in GLSL. The `ShadingRateInterlockOrderedEXT` and `ShadingRateInterlockUnorderedEXT` modes aren't implemented yet, because we don't support `SPV_NV_shading_rate` or `SPV_EXT_fragment_invocation_density` yet. HLSL and MSL were more interesting. They don't support this directly, but they do support marking resources as "rasterizer ordered," which does roughly the same thing. So this implementation scans all accesses inside the critical section and marks all storage resources found therein as rasterizer ordered. They also don't support the fine-grained controls on pixel- vs. sample-level interlock and disabling ordering guarantees that GLSL and SPIR-V do, but that's OK. "Unordered" here merely means the order is undefined; that it just so happens to be the same as rasterizer order is immaterial. As for pixel- vs. sample-level interlock, Vulkan explicitly states: > With sample shading enabled, [the `PixelInterlockOrderedEXT` and > `PixelInterlockUnorderedEXT`] execution modes are treated like > `SampleInterlockOrderedEXT` or `SampleInterlockUnorderedEXT` > respectively. and: > If [the `SampleInterlockOrderedEXT` or `SampleInterlockUnorderedEXT`] > execution modes are used in single-sample mode they are treated like > `PixelInterlockOrderedEXT` or `PixelInterlockUnorderedEXT` > respectively. So this will DTRT for MoltenVK and gfx-rs, at least. MSL additionally supports multiple raster order groups; resources that are not accessed together can be placed in different ROGs to allow them to be synchronized separately. A more sophisticated analysis might be able to place resources optimally, but that's outside the scope of this change. For now, we assign all resources to group 0, which should do for our purposes. `glslang` doesn't support the `RasterizerOrdered` UAVs this implementation produces for HLSL, so the test case needs `fxc.exe`. It also insists on GLSL 4.50 for `GL_ARB_fragment_shader_interlock`, even though the spec says it needs either 4.20 or `GL_ARB_shader_image_load_store`; and it doesn't support the `GL_NV_fragment_shader_interlock` extension at all. So I haven't been able to test those code paths. Fixes #1002.
2019-08-04 05:07:20 +00:00
: compiler(compiler_)
{
call_stack.push_back(entry_point_id);
Support the SPV_EXT_fragment_shader_interlock extension. This was straightforward to implement in GLSL. The `ShadingRateInterlockOrderedEXT` and `ShadingRateInterlockUnorderedEXT` modes aren't implemented yet, because we don't support `SPV_NV_shading_rate` or `SPV_EXT_fragment_invocation_density` yet. HLSL and MSL were more interesting. They don't support this directly, but they do support marking resources as "rasterizer ordered," which does roughly the same thing. So this implementation scans all accesses inside the critical section and marks all storage resources found therein as rasterizer ordered. They also don't support the fine-grained controls on pixel- vs. sample-level interlock and disabling ordering guarantees that GLSL and SPIR-V do, but that's OK. "Unordered" here merely means the order is undefined; that it just so happens to be the same as rasterizer order is immaterial. As for pixel- vs. sample-level interlock, Vulkan explicitly states: > With sample shading enabled, [the `PixelInterlockOrderedEXT` and > `PixelInterlockUnorderedEXT`] execution modes are treated like > `SampleInterlockOrderedEXT` or `SampleInterlockUnorderedEXT` > respectively. and: > If [the `SampleInterlockOrderedEXT` or `SampleInterlockUnorderedEXT`] > execution modes are used in single-sample mode they are treated like > `PixelInterlockOrderedEXT` or `PixelInterlockUnorderedEXT` > respectively. So this will DTRT for MoltenVK and gfx-rs, at least. MSL additionally supports multiple raster order groups; resources that are not accessed together can be placed in different ROGs to allow them to be synchronized separately. A more sophisticated analysis might be able to place resources optimally, but that's outside the scope of this change. For now, we assign all resources to group 0, which should do for our purposes. `glslang` doesn't support the `RasterizerOrdered` UAVs this implementation produces for HLSL, so the test case needs `fxc.exe`. It also insists on GLSL 4.50 for `GL_ARB_fragment_shader_interlock`, even though the spec says it needs either 4.20 or `GL_ARB_shader_image_load_store`; and it doesn't support the `GL_NV_fragment_shader_interlock` extension at all. So I haven't been able to test those code paths. Fixes #1002.
2019-08-04 05:07:20 +00:00
}
bool handle(spv::Op op, const uint32_t *args, uint32_t length) override;
bool begin_function_scope(const uint32_t *args, uint32_t length) override;
bool end_function_scope(const uint32_t *args, uint32_t length) override;
Support the SPV_EXT_fragment_shader_interlock extension. This was straightforward to implement in GLSL. The `ShadingRateInterlockOrderedEXT` and `ShadingRateInterlockUnorderedEXT` modes aren't implemented yet, because we don't support `SPV_NV_shading_rate` or `SPV_EXT_fragment_invocation_density` yet. HLSL and MSL were more interesting. They don't support this directly, but they do support marking resources as "rasterizer ordered," which does roughly the same thing. So this implementation scans all accesses inside the critical section and marks all storage resources found therein as rasterizer ordered. They also don't support the fine-grained controls on pixel- vs. sample-level interlock and disabling ordering guarantees that GLSL and SPIR-V do, but that's OK. "Unordered" here merely means the order is undefined; that it just so happens to be the same as rasterizer order is immaterial. As for pixel- vs. sample-level interlock, Vulkan explicitly states: > With sample shading enabled, [the `PixelInterlockOrderedEXT` and > `PixelInterlockUnorderedEXT`] execution modes are treated like > `SampleInterlockOrderedEXT` or `SampleInterlockUnorderedEXT` > respectively. and: > If [the `SampleInterlockOrderedEXT` or `SampleInterlockUnorderedEXT`] > execution modes are used in single-sample mode they are treated like > `PixelInterlockOrderedEXT` or `PixelInterlockUnorderedEXT` > respectively. So this will DTRT for MoltenVK and gfx-rs, at least. MSL additionally supports multiple raster order groups; resources that are not accessed together can be placed in different ROGs to allow them to be synchronized separately. A more sophisticated analysis might be able to place resources optimally, but that's outside the scope of this change. For now, we assign all resources to group 0, which should do for our purposes. `glslang` doesn't support the `RasterizerOrdered` UAVs this implementation produces for HLSL, so the test case needs `fxc.exe`. It also insists on GLSL 4.50 for `GL_ARB_fragment_shader_interlock`, even though the spec says it needs either 4.20 or `GL_ARB_shader_image_load_store`; and it doesn't support the `GL_NV_fragment_shader_interlock` extension at all. So I haven't been able to test those code paths. Fixes #1002.
2019-08-04 05:07:20 +00:00
Compiler &compiler;
bool in_crit_sec = false;
uint32_t interlock_function_id = 0;
bool split_function_case = false;
bool control_flow_interlock = false;
bool use_critical_section = false;
bool call_stack_is_interlocked = false;
SmallVector<uint32_t> call_stack;
void access_potential_resource(uint32_t id);
};
struct InterlockedResourceAccessPrepassHandler : OpcodeHandler
{
InterlockedResourceAccessPrepassHandler(Compiler &compiler_, uint32_t entry_point_id)
: compiler(compiler_)
{
call_stack.push_back(entry_point_id);
}
void rearm_current_block(const SPIRBlock &block) override;
bool handle(spv::Op op, const uint32_t *args, uint32_t length) override;
bool begin_function_scope(const uint32_t *args, uint32_t length) override;
bool end_function_scope(const uint32_t *args, uint32_t length) override;
Compiler &compiler;
uint32_t interlock_function_id = 0;
uint32_t current_block_id = 0;
bool split_function_case = false;
bool control_flow_interlock = false;
SmallVector<uint32_t> call_stack;
Support the SPV_EXT_fragment_shader_interlock extension. This was straightforward to implement in GLSL. The `ShadingRateInterlockOrderedEXT` and `ShadingRateInterlockUnorderedEXT` modes aren't implemented yet, because we don't support `SPV_NV_shading_rate` or `SPV_EXT_fragment_invocation_density` yet. HLSL and MSL were more interesting. They don't support this directly, but they do support marking resources as "rasterizer ordered," which does roughly the same thing. So this implementation scans all accesses inside the critical section and marks all storage resources found therein as rasterizer ordered. They also don't support the fine-grained controls on pixel- vs. sample-level interlock and disabling ordering guarantees that GLSL and SPIR-V do, but that's OK. "Unordered" here merely means the order is undefined; that it just so happens to be the same as rasterizer order is immaterial. As for pixel- vs. sample-level interlock, Vulkan explicitly states: > With sample shading enabled, [the `PixelInterlockOrderedEXT` and > `PixelInterlockUnorderedEXT`] execution modes are treated like > `SampleInterlockOrderedEXT` or `SampleInterlockUnorderedEXT` > respectively. and: > If [the `SampleInterlockOrderedEXT` or `SampleInterlockUnorderedEXT`] > execution modes are used in single-sample mode they are treated like > `PixelInterlockOrderedEXT` or `PixelInterlockUnorderedEXT` > respectively. So this will DTRT for MoltenVK and gfx-rs, at least. MSL additionally supports multiple raster order groups; resources that are not accessed together can be placed in different ROGs to allow them to be synchronized separately. A more sophisticated analysis might be able to place resources optimally, but that's outside the scope of this change. For now, we assign all resources to group 0, which should do for our purposes. `glslang` doesn't support the `RasterizerOrdered` UAVs this implementation produces for HLSL, so the test case needs `fxc.exe`. It also insists on GLSL 4.50 for `GL_ARB_fragment_shader_interlock`, even though the spec says it needs either 4.20 or `GL_ARB_shader_image_load_store`; and it doesn't support the `GL_NV_fragment_shader_interlock` extension at all. So I haven't been able to test those code paths. Fixes #1002.
2019-08-04 05:07:20 +00:00
};
void analyze_interlocked_resource_usage();
// The set of all resources written while inside the critical section, if present.
std::unordered_set<uint32_t> interlocked_resources;
bool interlocked_is_complex = false;
Support the SPV_EXT_fragment_shader_interlock extension. This was straightforward to implement in GLSL. The `ShadingRateInterlockOrderedEXT` and `ShadingRateInterlockUnorderedEXT` modes aren't implemented yet, because we don't support `SPV_NV_shading_rate` or `SPV_EXT_fragment_invocation_density` yet. HLSL and MSL were more interesting. They don't support this directly, but they do support marking resources as "rasterizer ordered," which does roughly the same thing. So this implementation scans all accesses inside the critical section and marks all storage resources found therein as rasterizer ordered. They also don't support the fine-grained controls on pixel- vs. sample-level interlock and disabling ordering guarantees that GLSL and SPIR-V do, but that's OK. "Unordered" here merely means the order is undefined; that it just so happens to be the same as rasterizer order is immaterial. As for pixel- vs. sample-level interlock, Vulkan explicitly states: > With sample shading enabled, [the `PixelInterlockOrderedEXT` and > `PixelInterlockUnorderedEXT`] execution modes are treated like > `SampleInterlockOrderedEXT` or `SampleInterlockUnorderedEXT` > respectively. and: > If [the `SampleInterlockOrderedEXT` or `SampleInterlockUnorderedEXT`] > execution modes are used in single-sample mode they are treated like > `PixelInterlockOrderedEXT` or `PixelInterlockUnorderedEXT` > respectively. So this will DTRT for MoltenVK and gfx-rs, at least. MSL additionally supports multiple raster order groups; resources that are not accessed together can be placed in different ROGs to allow them to be synchronized separately. A more sophisticated analysis might be able to place resources optimally, but that's outside the scope of this change. For now, we assign all resources to group 0, which should do for our purposes. `glslang` doesn't support the `RasterizerOrdered` UAVs this implementation produces for HLSL, so the test case needs `fxc.exe`. It also insists on GLSL 4.50 for `GL_ARB_fragment_shader_interlock`, even though the spec says it needs either 4.20 or `GL_ARB_shader_image_load_store`; and it doesn't support the `GL_NV_fragment_shader_interlock` extension at all. So I haven't been able to test those code paths. Fixes #1002.
2019-08-04 05:07:20 +00:00
2017-08-03 12:32:07 +00:00
void make_constant_null(uint32_t id, uint32_t type);
std::unordered_map<uint32_t, std::string> declared_block_names;
bool instruction_to_result_type(uint32_t &result_type, uint32_t &result_id, spv::Op op, const uint32_t *args,
uint32_t length);
2018-06-18 16:30:16 +00:00
Bitset combined_decoration_for_member(const SPIRType &type, uint32_t index) const;
2018-06-20 17:25:38 +00:00
static bool is_desktop_only_format(spv::ImageFormat format);
2018-06-18 16:30:16 +00:00
bool is_depth_image(const SPIRType &type, uint32_t id) const;
void set_extended_decoration(uint32_t id, ExtendedDecorations decoration, uint32_t value = 0);
uint32_t get_extended_decoration(uint32_t id, ExtendedDecorations decoration) const;
bool has_extended_decoration(uint32_t id, ExtendedDecorations decoration) const;
void unset_extended_decoration(uint32_t id, ExtendedDecorations decoration);
2019-01-17 10:29:50 +00:00
void set_extended_member_decoration(uint32_t type, uint32_t index, ExtendedDecorations decoration,
uint32_t value = 0);
uint32_t get_extended_member_decoration(uint32_t type, uint32_t index, ExtendedDecorations decoration) const;
bool has_extended_member_decoration(uint32_t type, uint32_t index, ExtendedDecorations decoration) const;
void unset_extended_member_decoration(uint32_t type, uint32_t index, ExtendedDecorations decoration);
bool check_internal_recursion(const SPIRType &type, std::unordered_set<uint32_t> &checked_ids);
bool type_contains_recursion(const SPIRType &type);
bool type_is_array_of_pointers(const SPIRType &type) const;
bool type_is_block_like(const SPIRType &type) const;
bool type_is_top_level_block(const SPIRType &type) const;
bool type_is_opaque_value(const SPIRType &type) const;
bool reflection_ssbo_instance_name_is_significant() const;
std::string get_remapped_declared_block_name(uint32_t id, bool fallback_prefer_instance_name) const;
bool flush_phi_required(BlockID from, BlockID to) const;
uint32_t evaluate_spec_constant_u32(const SPIRConstantOp &spec) const;
uint32_t evaluate_constant_u32(uint32_t id) const;
bool is_vertex_like_shader() const;
// Get the correct case list for the OpSwitch, since it can be either a
// 32 bit wide condition or a 64 bit, but the type is not embedded in the
// instruction itself.
const SmallVector<SPIRBlock::Case> &get_case_list(const SPIRBlock &block) const;
private:
// Used only to implement the old deprecated get_entry_point() interface.
const SPIREntryPoint &get_first_entry_point(const std::string &name) const;
SPIREntryPoint &get_first_entry_point(const std::string &name);
};
} // namespace SPIRV_CROSS_NAMESPACE
2016-03-02 17:09:16 +00:00
#endif