SPIRV-Tools/source/fuzz/fuzzer_pass_donate_modules.cpp
Alastair Donaldson 9c4481419e
spirv-fuzz: Allow inapplicable transformations to be ignored (#4407)
spirv-fuzz features transformations that should be applicable by
construction. Assertions are used to detect when such transformations
turn out to be inapplicable. Failures of such assertions indicate bugs
in the fuzzer. However, when using the fuzzer at scale (e.g. in
ClusterFuzz) reports of these assertion failures create noise, and
cause the fuzzer to exit early. This change adds an option whereby
inapplicable transformations can be ignored. This reduces noise and
allows fuzzing to continue even when a transformation that should be
applicable but is not has been erroneously created.
2021-07-28 22:59:37 +01:00

1216 lines
54 KiB
C++

// Copyright (c) 2019 Google LLC
//
// 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.
#include "source/fuzz/fuzzer_pass_donate_modules.h"
#include <map>
#include <queue>
#include <set>
#include "source/fuzz/call_graph.h"
#include "source/fuzz/instruction_message.h"
#include "source/fuzz/transformation_add_constant_boolean.h"
#include "source/fuzz/transformation_add_constant_composite.h"
#include "source/fuzz/transformation_add_constant_null.h"
#include "source/fuzz/transformation_add_constant_scalar.h"
#include "source/fuzz/transformation_add_function.h"
#include "source/fuzz/transformation_add_global_undef.h"
#include "source/fuzz/transformation_add_global_variable.h"
#include "source/fuzz/transformation_add_spec_constant_op.h"
#include "source/fuzz/transformation_add_type_array.h"
#include "source/fuzz/transformation_add_type_boolean.h"
#include "source/fuzz/transformation_add_type_float.h"
#include "source/fuzz/transformation_add_type_function.h"
#include "source/fuzz/transformation_add_type_int.h"
#include "source/fuzz/transformation_add_type_matrix.h"
#include "source/fuzz/transformation_add_type_pointer.h"
#include "source/fuzz/transformation_add_type_struct.h"
#include "source/fuzz/transformation_add_type_vector.h"
namespace spvtools {
namespace fuzz {
FuzzerPassDonateModules::FuzzerPassDonateModules(
opt::IRContext* ir_context, TransformationContext* transformation_context,
FuzzerContext* fuzzer_context,
protobufs::TransformationSequence* transformations,
bool ignore_inapplicable_transformations,
std::vector<fuzzerutil::ModuleSupplier> donor_suppliers)
: FuzzerPass(ir_context, transformation_context, fuzzer_context,
transformations, ignore_inapplicable_transformations),
donor_suppliers_(std::move(donor_suppliers)) {}
void FuzzerPassDonateModules::Apply() {
// If there are no donor suppliers, this fuzzer pass is a no-op.
if (donor_suppliers_.empty()) {
return;
}
// Donate at least one module, and probabilistically decide when to stop
// donating modules.
do {
// Choose a donor supplier at random, and get the module that it provides.
std::unique_ptr<opt::IRContext> donor_ir_context = donor_suppliers_.at(
GetFuzzerContext()->RandomIndex(donor_suppliers_))();
assert(donor_ir_context != nullptr && "Supplying of donor failed");
assert(
fuzzerutil::IsValid(donor_ir_context.get(),
GetTransformationContext()->GetValidatorOptions(),
fuzzerutil::kSilentMessageConsumer) &&
"The donor module must be valid");
// Donate the supplied module.
//
// Randomly decide whether to make the module livesafe (see
// FactFunctionIsLivesafe); doing so allows it to be used for live code
// injection but restricts its behaviour to allow this, and means that its
// functions cannot be transformed as if they were arbitrary dead code.
bool make_livesafe = GetFuzzerContext()->ChoosePercentage(
GetFuzzerContext()->ChanceOfMakingDonorLivesafe());
DonateSingleModule(donor_ir_context.get(), make_livesafe);
} while (GetFuzzerContext()->ChoosePercentage(
GetFuzzerContext()->GetChanceOfDonatingAdditionalModule()));
}
void FuzzerPassDonateModules::DonateSingleModule(
opt::IRContext* donor_ir_context, bool make_livesafe) {
// Check that the donated module has capabilities, supported by the recipient
// module.
for (const auto& capability_inst : donor_ir_context->capabilities()) {
auto capability =
static_cast<SpvCapability>(capability_inst.GetSingleWordInOperand(0));
if (!GetIRContext()->get_feature_mgr()->HasCapability(capability)) {
return;
}
}
// The ids used by the donor module may very well clash with ids defined in
// the recipient module. Furthermore, some instructions defined in the donor
// module will be equivalent to instructions defined in the recipient module,
// and it is not always legal to re-declare equivalent instructions. For
// example, OpTypeVoid cannot be declared twice.
//
// To handle this, we maintain a mapping from an id used in the donor module
// to the corresponding id that will be used by the donated code when it
// appears in the recipient module.
//
// This mapping is populated in two ways:
// (1) by mapping a donor instruction's result id to the id of some equivalent
// existing instruction in the recipient (e.g. this has to be done for
// OpTypeVoid)
// (2) by mapping a donor instruction's result id to a freshly chosen id that
// is guaranteed to be different from any id already used by the recipient
// (or from any id already chosen to handle a previous donor id)
std::map<uint32_t, uint32_t> original_id_to_donated_id;
HandleExternalInstructionImports(donor_ir_context,
&original_id_to_donated_id);
HandleTypesAndValues(donor_ir_context, &original_id_to_donated_id);
HandleFunctions(donor_ir_context, &original_id_to_donated_id, make_livesafe);
// TODO(https://github.com/KhronosGroup/SPIRV-Tools/issues/3115) Handle some
// kinds of decoration.
}
SpvStorageClass FuzzerPassDonateModules::AdaptStorageClass(
SpvStorageClass donor_storage_class) {
switch (donor_storage_class) {
case SpvStorageClassFunction:
case SpvStorageClassPrivate:
case SpvStorageClassWorkgroup:
// We leave these alone
return donor_storage_class;
case SpvStorageClassInput:
case SpvStorageClassOutput:
case SpvStorageClassUniform:
case SpvStorageClassUniformConstant:
case SpvStorageClassPushConstant:
case SpvStorageClassImage:
case SpvStorageClassStorageBuffer:
// We change these to Private
return SpvStorageClassPrivate;
default:
// Handle other cases on demand.
assert(false && "Currently unsupported storage class.");
return SpvStorageClassMax;
}
}
void FuzzerPassDonateModules::HandleExternalInstructionImports(
opt::IRContext* donor_ir_context,
std::map<uint32_t, uint32_t>* original_id_to_donated_id) {
// Consider every external instruction set import in the donor module.
for (auto& donor_import : donor_ir_context->module()->ext_inst_imports()) {
const auto& donor_import_name_words = donor_import.GetInOperand(0).words;
// Look for an identical import in the recipient module.
for (auto& existing_import : GetIRContext()->module()->ext_inst_imports()) {
const auto& existing_import_name_words =
existing_import.GetInOperand(0).words;
if (donor_import_name_words == existing_import_name_words) {
// A matching import has found. Map the result id for the donor import
// to the id of the existing import, so that when donor instructions
// rely on the import they will be rewritten to use the existing import.
original_id_to_donated_id->insert(
{donor_import.result_id(), existing_import.result_id()});
break;
}
}
// TODO(https://github.com/KhronosGroup/SPIRV-Tools/issues/3116): At present
// we do not handle donation of instruction imports, i.e. we do not allow
// the donor to import instruction sets that the recipient did not already
// import. It might be a good idea to allow this, but it requires some
// thought.
assert(original_id_to_donated_id->count(donor_import.result_id()) &&
"Donation of imports is not yet supported.");
}
}
void FuzzerPassDonateModules::HandleTypesAndValues(
opt::IRContext* donor_ir_context,
std::map<uint32_t, uint32_t>* original_id_to_donated_id) {
// Consider every type/global/constant/undef in the module.
for (auto& type_or_value : donor_ir_context->module()->types_values()) {
HandleTypeOrValue(type_or_value, original_id_to_donated_id);
}
}
void FuzzerPassDonateModules::HandleTypeOrValue(
const opt::Instruction& type_or_value,
std::map<uint32_t, uint32_t>* original_id_to_donated_id) {
// The type/value instruction generates a result id, and we need to associate
// the donor's result id with a new result id. That new result id will either
// be the id of some existing instruction, or a fresh id. This variable
// captures it.
uint32_t new_result_id;
// Decide how to handle each kind of instruction on a case-by-case basis.
//
// Because the donor module is required to be valid, when we encounter a
// type comprised of component types (e.g. an aggregate or pointer), we know
// that its component types will have been considered previously, and that
// |original_id_to_donated_id| will already contain an entry for them.
switch (type_or_value.opcode()) {
case SpvOpTypeImage:
case SpvOpTypeSampledImage:
case SpvOpTypeSampler:
// We do not donate types and variables that relate to images and
// samplers, so we skip these types and subsequently skip anything that
// depends on them.
return;
case SpvOpTypeVoid: {
// Void has to exist already in order for us to have an entry point.
// Get the existing id of void.
opt::analysis::Void void_type;
new_result_id = GetIRContext()->get_type_mgr()->GetId(&void_type);
assert(new_result_id &&
"The module being transformed will always have 'void' type "
"declared.");
} break;
case SpvOpTypeBool: {
// Bool cannot be declared multiple times, so use its existing id if
// present, or add a declaration of Bool with a fresh id if not.
opt::analysis::Bool bool_type;
auto bool_type_id = GetIRContext()->get_type_mgr()->GetId(&bool_type);
if (bool_type_id) {
new_result_id = bool_type_id;
} else {
new_result_id = GetFuzzerContext()->GetFreshId();
ApplyTransformation(TransformationAddTypeBoolean(new_result_id));
}
} break;
case SpvOpTypeInt: {
// Int cannot be declared multiple times with the same width and
// signedness, so check whether an existing identical Int type is
// present and use its id if so. Otherwise add a declaration of the
// Int type used by the donor, with a fresh id.
const uint32_t width = type_or_value.GetSingleWordInOperand(0);
const bool is_signed =
static_cast<bool>(type_or_value.GetSingleWordInOperand(1));
opt::analysis::Integer int_type(width, is_signed);
auto int_type_id = GetIRContext()->get_type_mgr()->GetId(&int_type);
if (int_type_id) {
new_result_id = int_type_id;
} else {
new_result_id = GetFuzzerContext()->GetFreshId();
ApplyTransformation(
TransformationAddTypeInt(new_result_id, width, is_signed));
}
} break;
case SpvOpTypeFloat: {
// Similar to SpvOpTypeInt.
const uint32_t width = type_or_value.GetSingleWordInOperand(0);
opt::analysis::Float float_type(width);
auto float_type_id = GetIRContext()->get_type_mgr()->GetId(&float_type);
if (float_type_id) {
new_result_id = float_type_id;
} else {
new_result_id = GetFuzzerContext()->GetFreshId();
ApplyTransformation(TransformationAddTypeFloat(new_result_id, width));
}
} break;
case SpvOpTypeVector: {
// It is not legal to have two Vector type declarations with identical
// element types and element counts, so check whether an existing
// identical Vector type is present and use its id if so. Otherwise add
// a declaration of the Vector type used by the donor, with a fresh id.
// When considering the vector's component type id, we look up the id
// use in the donor to find the id to which this has been remapped.
uint32_t component_type_id = original_id_to_donated_id->at(
type_or_value.GetSingleWordInOperand(0));
auto component_type =
GetIRContext()->get_type_mgr()->GetType(component_type_id);
assert(component_type && "The base type should be registered.");
auto component_count = type_or_value.GetSingleWordInOperand(1);
opt::analysis::Vector vector_type(component_type, component_count);
auto vector_type_id = GetIRContext()->get_type_mgr()->GetId(&vector_type);
if (vector_type_id) {
new_result_id = vector_type_id;
} else {
new_result_id = GetFuzzerContext()->GetFreshId();
ApplyTransformation(TransformationAddTypeVector(
new_result_id, component_type_id, component_count));
}
} break;
case SpvOpTypeMatrix: {
// Similar to SpvOpTypeVector.
uint32_t column_type_id = original_id_to_donated_id->at(
type_or_value.GetSingleWordInOperand(0));
auto column_type =
GetIRContext()->get_type_mgr()->GetType(column_type_id);
assert(column_type && column_type->AsVector() &&
"The column type should be a registered vector type.");
auto column_count = type_or_value.GetSingleWordInOperand(1);
opt::analysis::Matrix matrix_type(column_type, column_count);
auto matrix_type_id = GetIRContext()->get_type_mgr()->GetId(&matrix_type);
if (matrix_type_id) {
new_result_id = matrix_type_id;
} else {
new_result_id = GetFuzzerContext()->GetFreshId();
ApplyTransformation(TransformationAddTypeMatrix(
new_result_id, column_type_id, column_count));
}
} break;
case SpvOpTypeArray: {
// It is OK to have multiple structurally identical array types, so
// we go ahead and add a remapped version of the type declared by the
// donor.
uint32_t component_type_id = type_or_value.GetSingleWordInOperand(0);
if (!original_id_to_donated_id->count(component_type_id)) {
// We did not donate the component type of this array type, so we
// cannot donate the array type.
return;
}
new_result_id = GetFuzzerContext()->GetFreshId();
ApplyTransformation(TransformationAddTypeArray(
new_result_id, original_id_to_donated_id->at(component_type_id),
original_id_to_donated_id->at(
type_or_value.GetSingleWordInOperand(1))));
} break;
case SpvOpTypeRuntimeArray: {
// A runtime array is allowed as the final member of an SSBO. During
// donation we turn runtime arrays into fixed-size arrays. For dead
// code donations this is OK because the array is never indexed into at
// runtime, so it does not matter what its size is. For live-safe code,
// all accesses are made in-bounds, so this is also OK.
//
// The special OpArrayLength instruction, which works on runtime arrays,
// is rewritten to yield the fixed length that is used for the array.
uint32_t component_type_id = type_or_value.GetSingleWordInOperand(0);
if (!original_id_to_donated_id->count(component_type_id)) {
// We did not donate the component type of this runtime array type, so
// we cannot donate it as a fixed-size array.
return;
}
new_result_id = GetFuzzerContext()->GetFreshId();
ApplyTransformation(TransformationAddTypeArray(
new_result_id, original_id_to_donated_id->at(component_type_id),
FindOrCreateIntegerConstant(
{GetFuzzerContext()->GetRandomSizeForNewArray()}, 32, false,
false)));
} break;
case SpvOpTypeStruct: {
// Similar to SpvOpTypeArray.
std::vector<uint32_t> member_type_ids;
for (uint32_t i = 0; i < type_or_value.NumInOperands(); i++) {
auto component_type_id = type_or_value.GetSingleWordInOperand(i);
if (!original_id_to_donated_id->count(component_type_id)) {
// We did not donate every member type for this struct type, so we
// cannot donate the struct type.
return;
}
member_type_ids.push_back(
original_id_to_donated_id->at(component_type_id));
}
new_result_id = GetFuzzerContext()->GetFreshId();
ApplyTransformation(
TransformationAddTypeStruct(new_result_id, member_type_ids));
} break;
case SpvOpTypePointer: {
// Similar to SpvOpTypeArray.
uint32_t pointee_type_id = type_or_value.GetSingleWordInOperand(1);
if (!original_id_to_donated_id->count(pointee_type_id)) {
// We did not donate the pointee type for this pointer type, so we
// cannot donate the pointer type.
return;
}
new_result_id = GetFuzzerContext()->GetFreshId();
ApplyTransformation(TransformationAddTypePointer(
new_result_id,
AdaptStorageClass(static_cast<SpvStorageClass>(
type_or_value.GetSingleWordInOperand(0))),
original_id_to_donated_id->at(pointee_type_id)));
} break;
case SpvOpTypeFunction: {
// It is not OK to have multiple function types that use identical ids
// for their return and parameter types. We thus go through all
// existing function types to look for a match. We do not use the
// type manager here because we want to regard two function types that
// are structurally identical but that differ with respect to the
// actual ids used for pointer types as different.
//
// Example:
//
// %1 = OpTypeVoid
// %2 = OpTypeInt 32 0
// %3 = OpTypePointer Function %2
// %4 = OpTypePointer Function %2
// %5 = OpTypeFunction %1 %3
// %6 = OpTypeFunction %1 %4
//
// We regard %5 and %6 as distinct function types here, even though
// they both have the form "uint32* -> void"
std::vector<uint32_t> return_and_parameter_types;
for (uint32_t i = 0; i < type_or_value.NumInOperands(); i++) {
uint32_t return_or_parameter_type =
type_or_value.GetSingleWordInOperand(i);
if (!original_id_to_donated_id->count(return_or_parameter_type)) {
// We did not donate every return/parameter type for this function
// type, so we cannot donate the function type.
return;
}
return_and_parameter_types.push_back(
original_id_to_donated_id->at(return_or_parameter_type));
}
uint32_t existing_function_id = fuzzerutil::FindFunctionType(
GetIRContext(), return_and_parameter_types);
if (existing_function_id) {
new_result_id = existing_function_id;
} else {
// No match was found, so add a remapped version of the function type
// to the module, with a fresh id.
new_result_id = GetFuzzerContext()->GetFreshId();
std::vector<uint32_t> argument_type_ids;
for (uint32_t i = 1; i < type_or_value.NumInOperands(); i++) {
argument_type_ids.push_back(original_id_to_donated_id->at(
type_or_value.GetSingleWordInOperand(i)));
}
ApplyTransformation(TransformationAddTypeFunction(
new_result_id,
original_id_to_donated_id->at(
type_or_value.GetSingleWordInOperand(0)),
argument_type_ids));
}
} break;
case SpvOpSpecConstantOp: {
new_result_id = GetFuzzerContext()->GetFreshId();
auto type_id = original_id_to_donated_id->at(type_or_value.type_id());
auto opcode = static_cast<SpvOp>(type_or_value.GetSingleWordInOperand(0));
// Make sure we take into account |original_id_to_donated_id| when
// computing operands for OpSpecConstantOp.
opt::Instruction::OperandList operands;
for (uint32_t i = 1; i < type_or_value.NumInOperands(); ++i) {
const auto& operand = type_or_value.GetInOperand(i);
auto data =
operand.type == SPV_OPERAND_TYPE_ID
? opt::Operand::OperandData{original_id_to_donated_id->at(
operand.words[0])}
: operand.words;
operands.push_back({operand.type, std::move(data)});
}
ApplyTransformation(TransformationAddSpecConstantOp(
new_result_id, type_id, opcode, std::move(operands)));
} break;
case SpvOpSpecConstantTrue:
case SpvOpSpecConstantFalse:
case SpvOpConstantTrue:
case SpvOpConstantFalse: {
// It is OK to have duplicate definitions of True and False, so add
// these to the module, using a remapped Bool type.
new_result_id = GetFuzzerContext()->GetFreshId();
auto value = type_or_value.opcode() == SpvOpConstantTrue ||
type_or_value.opcode() == SpvOpSpecConstantTrue;
ApplyTransformation(
TransformationAddConstantBoolean(new_result_id, value, false));
} break;
case SpvOpSpecConstant:
case SpvOpConstant: {
// It is OK to have duplicate constant definitions, so add this to the
// module using a remapped result type.
new_result_id = GetFuzzerContext()->GetFreshId();
std::vector<uint32_t> data_words;
type_or_value.ForEachInOperand([&data_words](const uint32_t* in_operand) {
data_words.push_back(*in_operand);
});
ApplyTransformation(TransformationAddConstantScalar(
new_result_id, original_id_to_donated_id->at(type_or_value.type_id()),
data_words, false));
} break;
case SpvOpSpecConstantComposite:
case SpvOpConstantComposite: {
assert(original_id_to_donated_id->count(type_or_value.type_id()) &&
"Composite types for which it is possible to create a constant "
"should have been donated.");
// It is OK to have duplicate constant composite definitions, so add
// this to the module using remapped versions of all consituent ids and
// the result type.
new_result_id = GetFuzzerContext()->GetFreshId();
std::vector<uint32_t> constituent_ids;
type_or_value.ForEachInId([&constituent_ids, &original_id_to_donated_id](
const uint32_t* constituent_id) {
assert(original_id_to_donated_id->count(*constituent_id) &&
"The constants used to construct this composite should "
"have been donated.");
constituent_ids.push_back(
original_id_to_donated_id->at(*constituent_id));
});
ApplyTransformation(TransformationAddConstantComposite(
new_result_id, original_id_to_donated_id->at(type_or_value.type_id()),
constituent_ids, false));
} break;
case SpvOpConstantNull: {
if (!original_id_to_donated_id->count(type_or_value.type_id())) {
// We did not donate the type associated with this null constant, so
// we cannot donate the null constant.
return;
}
// It is fine to have multiple OpConstantNull instructions of the same
// type, so we just add this to the recipient module.
new_result_id = GetFuzzerContext()->GetFreshId();
ApplyTransformation(TransformationAddConstantNull(
new_result_id,
original_id_to_donated_id->at(type_or_value.type_id())));
} break;
case SpvOpVariable: {
if (!original_id_to_donated_id->count(type_or_value.type_id())) {
// We did not donate the pointer type associated with this variable,
// so we cannot donate the variable.
return;
}
// This is a global variable that could have one of various storage
// classes. However, we change all global variable pointer storage
// classes (such as Uniform, Input and Output) to private when donating
// pointer types, with the exception of the Workgroup storage class.
//
// Thus this variable's pointer type is guaranteed to have storage class
// Private or Workgroup.
//
// We add a global variable with either Private or Workgroup storage
// class, using remapped versions of the result type and initializer ids
// for the global variable in the donor.
//
// We regard the added variable as having an irrelevant value. This
// means that future passes can add stores to the variable in any
// way they wish, and pass them as pointer parameters to functions
// without worrying about whether their data might get modified.
new_result_id = GetFuzzerContext()->GetFreshId();
uint32_t remapped_pointer_type =
original_id_to_donated_id->at(type_or_value.type_id());
uint32_t initializer_id;
SpvStorageClass storage_class =
static_cast<SpvStorageClass>(type_or_value.GetSingleWordInOperand(
0)) == SpvStorageClassWorkgroup
? SpvStorageClassWorkgroup
: SpvStorageClassPrivate;
if (type_or_value.NumInOperands() == 1) {
// The variable did not have an initializer. Initialize it to zero
// if it has Private storage class (to limit problems associated with
// uninitialized data), and leave it uninitialized if it has Workgroup
// storage class (as Workgroup variables cannot have initializers).
// TODO(https://github.com/KhronosGroup/SPIRV-Tools/issues/3275): we
// could initialize Workgroup variables at the start of an entry
// point, and should do so if their uninitialized nature proves
// problematic.
initializer_id = storage_class == SpvStorageClassWorkgroup
? 0
: FindOrCreateZeroConstant(
fuzzerutil::GetPointeeTypeIdFromPointerType(
GetIRContext(), remapped_pointer_type),
false);
} else {
// The variable already had an initializer; use its remapped id.
initializer_id = original_id_to_donated_id->at(
type_or_value.GetSingleWordInOperand(1));
}
ApplyTransformation(
TransformationAddGlobalVariable(new_result_id, remapped_pointer_type,
storage_class, initializer_id, true));
} break;
case SpvOpUndef: {
if (!original_id_to_donated_id->count(type_or_value.type_id())) {
// We did not donate the type associated with this undef, so we cannot
// donate the undef.
return;
}
// It is fine to have multiple Undef instructions of the same type, so
// we just add this to the recipient module.
new_result_id = GetFuzzerContext()->GetFreshId();
ApplyTransformation(TransformationAddGlobalUndef(
new_result_id,
original_id_to_donated_id->at(type_or_value.type_id())));
} break;
default: {
assert(0 && "Unknown type/value.");
new_result_id = 0;
} break;
}
// Update the id mapping to associate the instruction's result id with its
// corresponding id in the recipient.
original_id_to_donated_id->insert({type_or_value.result_id(), new_result_id});
}
void FuzzerPassDonateModules::HandleFunctions(
opt::IRContext* donor_ir_context,
std::map<uint32_t, uint32_t>* original_id_to_donated_id,
bool make_livesafe) {
// Get the ids of functions in the donor module, topologically sorted
// according to the donor's call graph.
auto topological_order =
CallGraph(donor_ir_context).GetFunctionsInTopologicalOrder();
// Donate the functions in reverse topological order. This ensures that a
// function gets donated before any function that depends on it. This allows
// donation of the functions to be separated into a number of transformations,
// each adding one function, such that every prefix of transformations leaves
// the module valid.
for (auto function_id = topological_order.rbegin();
function_id != topological_order.rend(); ++function_id) {
// Find the function to be donated.
opt::Function* function_to_donate = nullptr;
for (auto& function : *donor_ir_context->module()) {
if (function.result_id() == *function_id) {
function_to_donate = &function;
break;
}
}
assert(function_to_donate && "Function to be donated was not found.");
if (!original_id_to_donated_id->count(
function_to_donate->DefInst().GetSingleWordInOperand(1))) {
// We were not able to donate this function's type, so we cannot donate
// the function.
continue;
}
// We will collect up protobuf messages representing the donor function's
// instructions here, and use them to create an AddFunction transformation.
std::vector<protobufs::Instruction> donated_instructions;
// This set tracks the ids of those instructions for which donation was
// completely skipped: neither the instruction nor a substitute for it was
// donated.
std::set<uint32_t> skipped_instructions;
// Consider every instruction of the donor function.
function_to_donate->ForEachInst(
[this, &donated_instructions, donor_ir_context,
&original_id_to_donated_id,
&skipped_instructions](const opt::Instruction* instruction) {
if (instruction->opcode() == SpvOpArrayLength) {
// We treat OpArrayLength specially.
HandleOpArrayLength(*instruction, original_id_to_donated_id,
&donated_instructions);
} else if (!CanDonateInstruction(donor_ir_context, *instruction,
*original_id_to_donated_id,
skipped_instructions)) {
// This is an instruction that we cannot directly donate.
HandleDifficultInstruction(*instruction, original_id_to_donated_id,
&donated_instructions,
&skipped_instructions);
} else {
PrepareInstructionForDonation(*instruction, donor_ir_context,
original_id_to_donated_id,
&donated_instructions);
}
});
// If |make_livesafe| is true, try to add the function in a livesafe manner.
// Otherwise (if |make_lifesafe| is false or an attempt to make the function
// livesafe has failed), add the function in a non-livesafe manner.
if (!make_livesafe ||
!MaybeAddLivesafeFunction(*function_to_donate, donor_ir_context,
*original_id_to_donated_id,
donated_instructions)) {
ApplyTransformation(TransformationAddFunction(donated_instructions));
}
}
}
bool FuzzerPassDonateModules::CanDonateInstruction(
opt::IRContext* donor_ir_context, const opt::Instruction& instruction,
const std::map<uint32_t, uint32_t>& original_id_to_donated_id,
const std::set<uint32_t>& skipped_instructions) const {
if (instruction.type_id() &&
!original_id_to_donated_id.count(instruction.type_id())) {
// We could not donate the result type of this instruction, so we cannot
// donate the instruction.
return false;
}
// Now consider instructions we specifically want to skip because we do not
// yet support them.
switch (instruction.opcode()) {
case SpvOpAtomicLoad:
case SpvOpAtomicStore:
case SpvOpAtomicExchange:
case SpvOpAtomicCompareExchange:
case SpvOpAtomicCompareExchangeWeak:
case SpvOpAtomicIIncrement:
case SpvOpAtomicIDecrement:
case SpvOpAtomicIAdd:
case SpvOpAtomicISub:
case SpvOpAtomicSMin:
case SpvOpAtomicUMin:
case SpvOpAtomicSMax:
case SpvOpAtomicUMax:
case SpvOpAtomicAnd:
case SpvOpAtomicOr:
case SpvOpAtomicXor:
// We conservatively ignore all atomic instructions at present.
// TODO(https://github.com/KhronosGroup/SPIRV-Tools/issues/3276): Consider
// being less conservative here.
case SpvOpImageSampleImplicitLod:
case SpvOpImageSampleExplicitLod:
case SpvOpImageSampleDrefImplicitLod:
case SpvOpImageSampleDrefExplicitLod:
case SpvOpImageSampleProjImplicitLod:
case SpvOpImageSampleProjExplicitLod:
case SpvOpImageSampleProjDrefImplicitLod:
case SpvOpImageSampleProjDrefExplicitLod:
case SpvOpImageFetch:
case SpvOpImageGather:
case SpvOpImageDrefGather:
case SpvOpImageRead:
case SpvOpImageWrite:
case SpvOpImageSparseSampleImplicitLod:
case SpvOpImageSparseSampleExplicitLod:
case SpvOpImageSparseSampleDrefImplicitLod:
case SpvOpImageSparseSampleDrefExplicitLod:
case SpvOpImageSparseSampleProjImplicitLod:
case SpvOpImageSparseSampleProjExplicitLod:
case SpvOpImageSparseSampleProjDrefImplicitLod:
case SpvOpImageSparseSampleProjDrefExplicitLod:
case SpvOpImageSparseFetch:
case SpvOpImageSparseGather:
case SpvOpImageSparseDrefGather:
case SpvOpImageSparseRead:
case SpvOpImageSampleFootprintNV:
case SpvOpImage:
case SpvOpImageQueryFormat:
case SpvOpImageQueryLevels:
case SpvOpImageQueryLod:
case SpvOpImageQueryOrder:
case SpvOpImageQuerySamples:
case SpvOpImageQuerySize:
case SpvOpImageQuerySizeLod:
case SpvOpSampledImage:
// We ignore all instructions related to accessing images, since we do not
// donate images.
return false;
case SpvOpLoad:
switch (donor_ir_context->get_def_use_mgr()
->GetDef(instruction.type_id())
->opcode()) {
case SpvOpTypeImage:
case SpvOpTypeSampledImage:
case SpvOpTypeSampler:
// Again, we ignore instructions that relate to accessing images.
return false;
default:
break;
}
default:
break;
}
// Examine each id input operand to the instruction. If it turns out that we
// have skipped any of these operands then we cannot donate the instruction.
bool result = true;
instruction.WhileEachInId(
[donor_ir_context, &original_id_to_donated_id, &result,
&skipped_instructions](const uint32_t* in_id) -> bool {
if (!original_id_to_donated_id.count(*in_id)) {
// We do not have a mapped result id for this id operand. That either
// means that it is a forward reference (which is OK), that we skipped
// the instruction that generated it (which is not OK), or that it is
// the id of a function or global value that we did not donate (which
// is not OK). We check for the latter two cases.
if (skipped_instructions.count(*in_id) ||
// A function or global value does not have an associated basic
// block.
!donor_ir_context->get_instr_block(*in_id)) {
result = false;
return false;
}
}
return true;
});
return result;
}
bool FuzzerPassDonateModules::IsBasicType(
const opt::Instruction& instruction) const {
switch (instruction.opcode()) {
case SpvOpTypeArray:
case SpvOpTypeBool:
case SpvOpTypeFloat:
case SpvOpTypeInt:
case SpvOpTypeMatrix:
case SpvOpTypeStruct:
case SpvOpTypeVector:
return true;
default:
return false;
}
}
void FuzzerPassDonateModules::HandleOpArrayLength(
const opt::Instruction& instruction,
std::map<uint32_t, uint32_t>* original_id_to_donated_id,
std::vector<protobufs::Instruction>* donated_instructions) const {
assert(instruction.opcode() == SpvOpArrayLength &&
"Precondition: instruction must be OpArrayLength.");
uint32_t donated_variable_id =
original_id_to_donated_id->at(instruction.GetSingleWordInOperand(0));
auto donated_variable_instruction =
GetIRContext()->get_def_use_mgr()->GetDef(donated_variable_id);
auto pointer_to_struct_instruction =
GetIRContext()->get_def_use_mgr()->GetDef(
donated_variable_instruction->type_id());
assert(pointer_to_struct_instruction->opcode() == SpvOpTypePointer &&
"Type of variable must be pointer.");
auto donated_struct_type_instruction =
GetIRContext()->get_def_use_mgr()->GetDef(
pointer_to_struct_instruction->GetSingleWordInOperand(1));
assert(donated_struct_type_instruction->opcode() == SpvOpTypeStruct &&
"Pointee type of pointer used by OpArrayLength must be struct.");
assert(donated_struct_type_instruction->NumInOperands() ==
instruction.GetSingleWordInOperand(1) + 1 &&
"OpArrayLength must refer to the final member of the given "
"struct.");
uint32_t fixed_size_array_type_id =
donated_struct_type_instruction->GetSingleWordInOperand(
donated_struct_type_instruction->NumInOperands() - 1);
auto fixed_size_array_type_instruction =
GetIRContext()->get_def_use_mgr()->GetDef(fixed_size_array_type_id);
assert(fixed_size_array_type_instruction->opcode() == SpvOpTypeArray &&
"The donated array type must be fixed-size.");
auto array_size_id =
fixed_size_array_type_instruction->GetSingleWordInOperand(1);
if (instruction.result_id() &&
!original_id_to_donated_id->count(instruction.result_id())) {
original_id_to_donated_id->insert(
{instruction.result_id(), GetFuzzerContext()->GetFreshId()});
}
donated_instructions->push_back(MakeInstructionMessage(
SpvOpCopyObject, original_id_to_donated_id->at(instruction.type_id()),
original_id_to_donated_id->at(instruction.result_id()),
opt::Instruction::OperandList({{SPV_OPERAND_TYPE_ID, {array_size_id}}})));
}
void FuzzerPassDonateModules::HandleDifficultInstruction(
const opt::Instruction& instruction,
std::map<uint32_t, uint32_t>* original_id_to_donated_id,
std::vector<protobufs::Instruction>* donated_instructions,
std::set<uint32_t>* skipped_instructions) {
if (!instruction.result_id()) {
// It does not generate a result id, so it can be ignored.
return;
}
if (!original_id_to_donated_id->count(instruction.type_id())) {
// We cannot handle this instruction's result type, so we need to skip it
// all together.
skipped_instructions->insert(instruction.result_id());
return;
}
// We now attempt to replace the instruction with an OpCopyObject.
// TODO(https://github.com/KhronosGroup/SPIRV-Tools/issues/3278): We could do
// something more refined here - we could check which operands to the
// instruction could not be donated and replace those operands with
// references to other ids (such as constants), so that we still get an
// instruction with the opcode and easy-to-handle operands of the donor
// instruction.
auto remapped_type_id = original_id_to_donated_id->at(instruction.type_id());
if (!IsBasicType(
*GetIRContext()->get_def_use_mgr()->GetDef(remapped_type_id))) {
// The instruction has a non-basic result type, so we cannot replace it with
// an object copy of a constant. We thus skip it completely.
// TODO(https://github.com/KhronosGroup/SPIRV-Tools/issues/3279): We could
// instead look for an available id of the right type and generate an
// OpCopyObject of that id.
skipped_instructions->insert(instruction.result_id());
return;
}
// We are going to add an OpCopyObject instruction. Add a mapping for the
// result id of the original instruction if does not already exist (it may
// exist in the case that it has been forward-referenced).
if (!original_id_to_donated_id->count(instruction.result_id())) {
original_id_to_donated_id->insert(
{instruction.result_id(), GetFuzzerContext()->GetFreshId()});
}
// We find or add a zero constant to the receiving module for the type in
// question, and add an OpCopyObject instruction that copies this zero.
//
// We mark the constant as irrelevant so that we can replace it with a
// more interesting value later.
auto zero_constant = FindOrCreateZeroConstant(remapped_type_id, true);
donated_instructions->push_back(MakeInstructionMessage(
SpvOpCopyObject, remapped_type_id,
original_id_to_donated_id->at(instruction.result_id()),
opt::Instruction::OperandList({{SPV_OPERAND_TYPE_ID, {zero_constant}}})));
}
void FuzzerPassDonateModules::PrepareInstructionForDonation(
const opt::Instruction& instruction, opt::IRContext* donor_ir_context,
std::map<uint32_t, uint32_t>* original_id_to_donated_id,
std::vector<protobufs::Instruction>* donated_instructions) {
// Get the instruction's input operands into donation-ready form,
// remapping any id uses in the process.
opt::Instruction::OperandList input_operands;
// Consider each input operand in turn.
for (uint32_t in_operand_index = 0;
in_operand_index < instruction.NumInOperands(); in_operand_index++) {
std::vector<uint32_t> operand_data;
const opt::Operand& in_operand = instruction.GetInOperand(in_operand_index);
// Check whether this operand is an id.
if (spvIsIdType(in_operand.type)) {
// This is an id operand - it consists of a single word of data,
// which needs to be remapped so that it is replaced with the
// donated form of the id.
auto operand_id = in_operand.words[0];
if (!original_id_to_donated_id->count(operand_id)) {
// This is a forward reference. We will choose a corresponding
// donor id for the referenced id and update the mapping to
// reflect it.
// Keep release compilers happy because |donor_ir_context| is only used
// in this assertion.
(void)(donor_ir_context);
assert((donor_ir_context->get_def_use_mgr()
->GetDef(operand_id)
->opcode() == SpvOpLabel ||
instruction.opcode() == SpvOpPhi) &&
"Unsupported forward reference.");
original_id_to_donated_id->insert(
{operand_id, GetFuzzerContext()->GetFreshId()});
}
operand_data.push_back(original_id_to_donated_id->at(operand_id));
} else {
// For non-id operands, we just add each of the data words.
for (auto word : in_operand.words) {
operand_data.push_back(word);
}
}
input_operands.push_back({in_operand.type, operand_data});
}
if (instruction.opcode() == SpvOpVariable &&
instruction.NumInOperands() == 1) {
// This is an uninitialized local variable. Initialize it to zero.
input_operands.push_back(
{SPV_OPERAND_TYPE_ID,
{FindOrCreateZeroConstant(
fuzzerutil::GetPointeeTypeIdFromPointerType(
GetIRContext(),
original_id_to_donated_id->at(instruction.type_id())),
false)}});
}
if (instruction.result_id() &&
!original_id_to_donated_id->count(instruction.result_id())) {
original_id_to_donated_id->insert(
{instruction.result_id(), GetFuzzerContext()->GetFreshId()});
}
// Remap the result type and result id (if present) of the
// instruction, and turn it into a protobuf message.
donated_instructions->push_back(MakeInstructionMessage(
instruction.opcode(),
instruction.type_id()
? original_id_to_donated_id->at(instruction.type_id())
: 0,
instruction.result_id()
? original_id_to_donated_id->at(instruction.result_id())
: 0,
input_operands));
}
bool FuzzerPassDonateModules::CreateLoopLimiterInfo(
opt::IRContext* donor_ir_context, const opt::BasicBlock& loop_header,
const std::map<uint32_t, uint32_t>& original_id_to_donated_id,
protobufs::LoopLimiterInfo* out) {
assert(loop_header.IsLoopHeader() && "|loop_header| is not a loop header");
// Grab the loop header's id, mapped to its donated value.
out->set_loop_header_id(original_id_to_donated_id.at(loop_header.id()));
// Get fresh ids that will be used to load the loop limiter, increment
// it, compare it with the loop limit, and an id for a new block that
// will contain the loop's original terminator.
out->set_load_id(GetFuzzerContext()->GetFreshId());
out->set_increment_id(GetFuzzerContext()->GetFreshId());
out->set_compare_id(GetFuzzerContext()->GetFreshId());
out->set_logical_op_id(GetFuzzerContext()->GetFreshId());
// We are creating a branch from the back-edge block to the merge block. Thus,
// if merge block has any OpPhi instructions, we might need to adjust
// them.
// Note that the loop might have an unreachable back-edge block. This means
// that the loop can't iterate, so we don't need to adjust anything.
const auto back_edge_block_id = TransformationAddFunction::GetBackEdgeBlockId(
donor_ir_context, loop_header.id());
if (!back_edge_block_id) {
return true;
}
auto* back_edge_block = donor_ir_context->cfg()->block(back_edge_block_id);
assert(back_edge_block && "|back_edge_block_id| is invalid");
const auto* merge_block =
donor_ir_context->cfg()->block(loop_header.MergeBlockId());
assert(merge_block && "Loop header has invalid merge block id");
// We don't need to adjust anything if there is already a branch from
// the back-edge block to the merge block.
if (back_edge_block->IsSuccessor(merge_block)) {
return true;
}
// Adjust OpPhi instructions in the |merge_block|.
for (const auto& inst : *merge_block) {
if (inst.opcode() != SpvOpPhi) {
break;
}
// There is no simple way to ensure that a chosen operand for the OpPhi
// instruction will never cause any problems (e.g. if we choose an
// integer id, it might have a zero value when we branch from the back
// edge block. This might cause a division by 0 later in the function.).
// Thus, we ignore possible problems and proceed as follows:
// - if any of the existing OpPhi operands dominates the back-edge
// block - use it
// - if OpPhi has a basic type (see IsBasicType method) - create
// a zero constant
// - otherwise, we can't add a livesafe function.
uint32_t suitable_operand_id = 0;
for (uint32_t i = 0; i < inst.NumInOperands(); i += 2) {
auto dependency_inst_id = inst.GetSingleWordInOperand(i);
if (fuzzerutil::IdIsAvailableBeforeInstruction(
donor_ir_context, back_edge_block->terminator(),
dependency_inst_id)) {
suitable_operand_id = original_id_to_donated_id.at(dependency_inst_id);
break;
}
}
if (suitable_operand_id == 0 &&
IsBasicType(
*donor_ir_context->get_def_use_mgr()->GetDef(inst.type_id()))) {
// We mark this constant as irrelevant so that we can replace it
// with more interesting value later.
suitable_operand_id = FindOrCreateZeroConstant(
original_id_to_donated_id.at(inst.type_id()), true);
}
if (suitable_operand_id == 0) {
return false;
}
out->add_phi_id(suitable_operand_id);
}
return true;
}
bool FuzzerPassDonateModules::MaybeAddLivesafeFunction(
const opt::Function& function_to_donate, opt::IRContext* donor_ir_context,
const std::map<uint32_t, uint32_t>& original_id_to_donated_id,
const std::vector<protobufs::Instruction>& donated_instructions) {
// Various types and constants must be in place for a function to be made
// live-safe. Add them if not already present.
FindOrCreateBoolType(); // Needed for comparisons
FindOrCreatePointerToIntegerType(
32, false, SpvStorageClassFunction); // Needed for adding loop limiters
FindOrCreateIntegerConstant({0}, 32, false,
false); // Needed for initializing loop limiters
FindOrCreateIntegerConstant({1}, 32, false,
false); // Needed for incrementing loop limiters
// Get a fresh id for the variable that will be used as a loop limiter.
const uint32_t loop_limiter_variable_id = GetFuzzerContext()->GetFreshId();
// Choose a random loop limit, and add the required constant to the
// module if not already there.
const uint32_t loop_limit = FindOrCreateIntegerConstant(
{GetFuzzerContext()->GetRandomLoopLimit()}, 32, false, false);
// Consider every loop header in the function to donate, and create a
// structure capturing the ids to be used for manipulating the loop
// limiter each time the loop is iterated.
std::vector<protobufs::LoopLimiterInfo> loop_limiters;
for (auto& block : function_to_donate) {
if (block.IsLoopHeader()) {
protobufs::LoopLimiterInfo loop_limiter;
if (!CreateLoopLimiterInfo(donor_ir_context, block,
original_id_to_donated_id, &loop_limiter)) {
return false;
}
loop_limiters.emplace_back(std::move(loop_limiter));
}
}
// Consider every access chain in the function to donate, and create a
// structure containing the ids necessary to clamp the access chain
// indices to be in-bounds.
std::vector<protobufs::AccessChainClampingInfo> access_chain_clamping_info;
for (auto& block : function_to_donate) {
for (auto& inst : block) {
switch (inst.opcode()) {
case SpvOpAccessChain:
case SpvOpInBoundsAccessChain: {
protobufs::AccessChainClampingInfo clamping_info;
clamping_info.set_access_chain_id(
original_id_to_donated_id.at(inst.result_id()));
auto base_object = donor_ir_context->get_def_use_mgr()->GetDef(
inst.GetSingleWordInOperand(0));
assert(base_object && "The base object must exist.");
auto pointer_type = donor_ir_context->get_def_use_mgr()->GetDef(
base_object->type_id());
assert(pointer_type && pointer_type->opcode() == SpvOpTypePointer &&
"The base object must have pointer type.");
auto should_be_composite_type =
donor_ir_context->get_def_use_mgr()->GetDef(
pointer_type->GetSingleWordInOperand(1));
// Walk the access chain, creating fresh ids to facilitate
// clamping each index. For simplicity we do this for every
// index, even though constant indices will not end up being
// clamped.
for (uint32_t index = 1; index < inst.NumInOperands(); index++) {
auto compare_and_select_ids =
clamping_info.add_compare_and_select_ids();
compare_and_select_ids->set_first(GetFuzzerContext()->GetFreshId());
compare_and_select_ids->set_second(
GetFuzzerContext()->GetFreshId());
// Get the bound for the component being indexed into.
uint32_t bound;
if (should_be_composite_type->opcode() == SpvOpTypeRuntimeArray) {
// The donor is indexing into a runtime array. We do not
// donate runtime arrays. Instead, we donate a corresponding
// fixed-size array for every runtime array. We should thus
// find that donor composite type's result id maps to a fixed-
// size array.
auto fixed_size_array_type =
GetIRContext()->get_def_use_mgr()->GetDef(
original_id_to_donated_id.at(
should_be_composite_type->result_id()));
assert(fixed_size_array_type->opcode() == SpvOpTypeArray &&
"A runtime array type in the donor should have been "
"replaced by a fixed-sized array in the recipient.");
// The size of this fixed-size array is a suitable bound.
bound = fuzzerutil::GetBoundForCompositeIndex(
*fixed_size_array_type, GetIRContext());
} else {
bound = fuzzerutil::GetBoundForCompositeIndex(
*should_be_composite_type, donor_ir_context);
}
const uint32_t index_id = inst.GetSingleWordInOperand(index);
auto index_inst =
donor_ir_context->get_def_use_mgr()->GetDef(index_id);
auto index_type_inst = donor_ir_context->get_def_use_mgr()->GetDef(
index_inst->type_id());
assert(index_type_inst->opcode() == SpvOpTypeInt);
opt::analysis::Integer* index_int_type =
donor_ir_context->get_type_mgr()
->GetType(index_type_inst->result_id())
->AsInteger();
if (index_inst->opcode() != SpvOpConstant) {
// We will have to clamp this index, so we need a constant
// whose value is one less than the bound, to compare
// against and to use as the clamped value.
FindOrCreateIntegerConstant({bound - 1}, 32,
index_int_type->IsSigned(), false);
}
should_be_composite_type =
TransformationAddFunction::FollowCompositeIndex(
donor_ir_context, *should_be_composite_type, index_id);
}
access_chain_clamping_info.push_back(clamping_info);
break;
}
default:
break;
}
}
}
// If |function_to_donate| has non-void return type and contains an
// OpKill/OpUnreachable instruction, then a value is needed in order to turn
// these into instructions of the form OpReturnValue %value_id.
uint32_t kill_unreachable_return_value_id = 0;
auto function_return_type_inst =
donor_ir_context->get_def_use_mgr()->GetDef(function_to_donate.type_id());
if (function_return_type_inst->opcode() != SpvOpTypeVoid &&
fuzzerutil::FunctionContainsOpKillOrUnreachable(function_to_donate)) {
kill_unreachable_return_value_id = FindOrCreateZeroConstant(
original_id_to_donated_id.at(function_return_type_inst->result_id()),
false);
}
// Try to add the function in a livesafe manner. This may fail due to edge
// cases, e.g. where adding loop limiters changes dominance such that the
// module becomes invalid. It would be ideal to handle all such edge cases,
// but as they are rare it is more pragmatic to bail out of making the
// function livesafe if the transformation's precondition fails to hold.
return MaybeApplyTransformation(TransformationAddFunction(
donated_instructions, loop_limiter_variable_id, loop_limit, loop_limiters,
kill_unreachable_return_value_id, access_chain_clamping_info));
}
} // namespace fuzz
} // namespace spvtools