Add fuzzer pass to copy objects (#2853)

A new fuzzer pass that randomly introduces OpCopyObject instructions
that make copies of ids, and uses the fact manager to record the fact
that an id %id is synonymous with an id generated by an OpCopyObject
applied to %id. (A future pass will exploit such synonym facts.)
This commit is contained in:
Alastair Donaldson 2019-09-11 23:45:20 +01:00 коммит произвёл GitHub
Родитель 67b87f22cf
Коммит ad7f2c5c4c
Не найден ключ, соответствующий данной подписи
Идентификатор ключа GPG: 4AEE18F83AFDEB23
9 изменённых файлов: 254 добавлений и 32 удалений

Просмотреть файл

@ -34,6 +34,7 @@ if(SPIRV_BUILD_FUZZER)
fuzzer_pass_add_dead_breaks.h
fuzzer_pass_add_dead_continues.h
fuzzer_pass_add_useful_constructs.h
fuzzer_pass_copy_objects.h
fuzzer_pass_obfuscate_constants.h
fuzzer_pass_permute_blocks.h
fuzzer_pass_split_blocks.h
@ -69,6 +70,7 @@ if(SPIRV_BUILD_FUZZER)
fuzzer_pass_add_dead_breaks.cpp
fuzzer_pass_add_dead_continues.cpp
fuzzer_pass_add_useful_constructs.cpp
fuzzer_pass_copy_objects.cpp
fuzzer_pass_obfuscate_constants.cpp
fuzzer_pass_permute_blocks.cpp
fuzzer_pass_split_blocks.cpp

Просмотреть файл

@ -22,6 +22,7 @@
#include "source/fuzz/fuzzer_pass_add_dead_breaks.h"
#include "source/fuzz/fuzzer_pass_add_dead_continues.h"
#include "source/fuzz/fuzzer_pass_add_useful_constructs.h"
#include "source/fuzz/fuzzer_pass_copy_objects.h"
#include "source/fuzz/fuzzer_pass_obfuscate_constants.h"
#include "source/fuzz/fuzzer_pass_permute_blocks.h"
#include "source/fuzz/fuzzer_pass_split_blocks.h"
@ -107,6 +108,9 @@ Fuzzer::FuzzerResultStatus Fuzzer::Run(
.Apply();
// Apply some semantics-preserving passes.
FuzzerPassCopyObjects(ir_context.get(), &fact_manager, &fuzzer_context,
transformation_sequence_out)
.Apply();
FuzzerPassSplitBlocks(ir_context.get(), &fact_manager, &fuzzer_context,
transformation_sequence_out)
.Apply();

Просмотреть файл

@ -26,6 +26,7 @@ namespace {
const uint32_t kDefaultChanceOfAddingDeadBreak = 20;
const uint32_t kDefaultChanceOfAddingDeadContinue = 20;
const uint32_t kDefaultChanceOfCopyingObject = 20;
const uint32_t kDefaultChanceOfMovingBlockDown = 25;
const uint32_t kDefaultChanceOfObfuscatingConstant = 20;
const uint32_t kDefaultChanceOfSplittingBlock = 20;
@ -48,6 +49,7 @@ FuzzerContext::FuzzerContext(RandomGenerator* random_generator,
next_fresh_id_(min_fresh_id),
chance_of_adding_dead_break_(kDefaultChanceOfAddingDeadBreak),
chance_of_adding_dead_continue_(kDefaultChanceOfAddingDeadContinue),
chance_of_copying_object_(kDefaultChanceOfCopyingObject),
chance_of_moving_block_down_(kDefaultChanceOfMovingBlockDown),
chance_of_obfuscating_constant_(kDefaultChanceOfObfuscatingConstant),
chance_of_splitting_block_(kDefaultChanceOfSplittingBlock),

Просмотреть файл

@ -61,6 +61,7 @@ class FuzzerContext {
uint32_t GetChanceOfAddingDeadContinue() {
return chance_of_adding_dead_continue_;
}
uint32_t GetChanceOfCopyingObject() { return chance_of_copying_object_; }
uint32_t GetChanceOfMovingBlockDown() { return chance_of_moving_block_down_; }
uint32_t GetChanceOfObfuscatingConstant() {
return chance_of_obfuscating_constant_;
@ -83,6 +84,7 @@ class FuzzerContext {
// Keep them in alphabetical order.
uint32_t chance_of_adding_dead_break_;
uint32_t chance_of_adding_dead_continue_;
uint32_t chance_of_copying_object_;
uint32_t chance_of_moving_block_down_;
uint32_t chance_of_obfuscating_constant_;
uint32_t chance_of_splitting_block_;

Просмотреть файл

@ -0,0 +1,148 @@
// 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_copy_objects.h"
#include "source/fuzz/transformation_copy_object.h"
namespace spvtools {
namespace fuzz {
FuzzerPassCopyObjects::FuzzerPassCopyObjects(
opt::IRContext* ir_context, FactManager* fact_manager,
FuzzerContext* fuzzer_context,
protobufs::TransformationSequence* transformations)
: FuzzerPass(ir_context, fact_manager, fuzzer_context, transformations) {}
FuzzerPassCopyObjects::~FuzzerPassCopyObjects() = default;
void FuzzerPassCopyObjects::Apply() {
// Consider every block in every function.
for (auto& function : *GetIRContext()->module()) {
for (auto& block : function) {
// We now consider every instruction in the block, randomly deciding
// whether to add an object copy before the instruction.
// In order to insert an object copy instruction, we need to be able to
// identify the instruction a copy should be inserted before. We do this
// by tracking a base instruction, which must generate a result id, and an
// offset (to allow us to identify instructions that do not generate
// result ids).
// The initial base instruction is the block label.
uint32_t base = block.id();
uint32_t offset = 0;
// Consider every instruction in the block.
for (auto inst_it = block.begin(); inst_it != block.end(); ++inst_it) {
if (inst_it->HasResultId()) {
// In the case that the instruction has a result id, we use the
// instruction as its own base, with zero offset.
base = inst_it->result_id();
offset = 0;
} else {
// The instruction does not have a result id, so we need to identify
// it via the latest instruction that did have a result id (base), and
// an incremented offset.
offset++;
}
// Check whether it is legitimate to insert a copy before this
// instruction.
if (!TransformationCopyObject::CanInsertCopyBefore(inst_it)) {
continue;
}
// Randomly decide whether to try inserting an object copy here.
if (!GetFuzzerContext()->ChoosePercentage(
GetFuzzerContext()->GetChanceOfCopyingObject())) {
continue;
}
// Populate list of potential instructions that can be copied.
// TODO(afd) The following is (relatively) simple, but may end up being
// prohibitively inefficient, as it walks the whole dominator tree for
// every copy that is added.
std::vector<opt::Instruction*> copyable_instructions;
// Consider all global declarations
for (auto& global : GetIRContext()->module()->types_values()) {
if (TransformationCopyObject::IsCopyable(GetIRContext(), &global)) {
copyable_instructions.push_back(&global);
}
}
// Consider all previous instructions in this block
for (auto prev_inst_it = block.begin(); prev_inst_it != inst_it;
++prev_inst_it) {
if (TransformationCopyObject::IsCopyable(GetIRContext(),
&*prev_inst_it)) {
copyable_instructions.push_back(&*prev_inst_it);
}
}
// Walk the dominator tree to consider all instructions from dominating
// blocks
auto dominator_analysis =
GetIRContext()->GetDominatorAnalysis(&function);
for (auto next_dominator =
dominator_analysis->ImmediateDominator(&block);
next_dominator != nullptr;
next_dominator =
dominator_analysis->ImmediateDominator(next_dominator)) {
for (auto& dominating_inst : *next_dominator) {
if (TransformationCopyObject::IsCopyable(GetIRContext(),
&dominating_inst)) {
copyable_instructions.push_back(&dominating_inst);
}
}
}
// At this point, |copyable_instructions| contains all the instructions
// we might think of copying.
if (!copyable_instructions.empty()) {
// Choose a copyable instruction at random, and create and apply an
// object copying transformation based on it.
uint32_t index =
GetFuzzerContext()->RandomIndex(copyable_instructions);
TransformationCopyObject transformation(
copyable_instructions[index]->result_id(), base, offset,
GetFuzzerContext()->GetFreshId());
assert(
transformation.IsApplicable(GetIRContext(), *GetFactManager()) &&
"This transformation should be applicable by construction.");
transformation.Apply(GetIRContext(), GetFactManager());
*GetTransformations()->add_transformation() =
transformation.ToMessage();
if (!inst_it->HasResultId()) {
// We have inserted a new instruction before the current
// instruction, and we are tracking the current id-less instruction
// via an offset (offset) from a previous instruction (base) that
// has an id. We increment |offset| to reflect the newly-inserted
// instruction.
//
// This is slightly preferable to the alternative of setting |base|
// to be the result id of the new instruction, since on replay we
// might end up eliminating this copy but keeping a subsequent copy.
offset++;
}
}
}
}
}
}
} // namespace fuzz
} // namespace spvtools

Просмотреть файл

@ -0,0 +1,38 @@
// 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.
#ifndef SOURCE_FUZZ_FUZZER_PASS_COPY_OBJECTS_H_
#define SOURCE_FUZZ_FUZZER_PASS_COPY_OBJECTS_H_
#include "source/fuzz/fuzzer_pass.h"
namespace spvtools {
namespace fuzz {
// A fuzzer pass for adding adding copies of objects to the module.
class FuzzerPassCopyObjects : public FuzzerPass {
public:
FuzzerPassCopyObjects(opt::IRContext* ir_context, FactManager* fact_manager,
FuzzerContext* fuzzer_context,
protobufs::TransformationSequence* transformations);
~FuzzerPassCopyObjects();
void Apply() override;
};
} // namespace fuzz
} // namespace spvtools
#endif // SOURCE_FUZZ_FUZZER_PASS_COPY_OBJECTS_H_

Просмотреть файл

@ -25,6 +25,7 @@
#include "transformation_add_type_float.h"
#include "transformation_add_type_int.h"
#include "transformation_add_type_pointer.h"
#include "transformation_copy_object.h"
#include "transformation_move_block_down.h"
#include "transformation_replace_boolean_constant_with_constant_binary.h"
#include "transformation_replace_constant_with_uniform.h"
@ -59,6 +60,8 @@ std::unique_ptr<Transformation> Transformation::FromMessage(
case protobufs::Transformation::TransformationCase::kAddTypePointer:
return MakeUnique<TransformationAddTypePointer>(
message.add_type_pointer());
case protobufs::Transformation::TransformationCase::kCopyObject:
return MakeUnique<TransformationCopyObject>(message.copy_object());
case protobufs::Transformation::TransformationCase::kMoveBlockDown:
return MakeUnique<TransformationMoveBlockDown>(message.move_block_down());
case protobufs::Transformation::TransformationCase::
@ -71,13 +74,12 @@ std::unique_ptr<Transformation> Transformation::FromMessage(
message.replace_constant_with_uniform());
case protobufs::Transformation::TransformationCase::kSplitBlock:
return MakeUnique<TransformationSplitBlock>(message.split_block());
default:
assert(message.transformation_case() ==
protobufs::Transformation::TRANSFORMATION_NOT_SET &&
"Unhandled transformation type.");
case protobufs::Transformation::TRANSFORMATION_NOT_SET:
assert(false && "An unset transformation was encountered.");
return nullptr;
}
assert(false && "Should be unreachable as all cases must be handled above.");
return nullptr;
}
} // namespace fuzz

Просмотреть файл

@ -46,18 +46,7 @@ bool TransformationCopyObject::IsApplicable(
if (!object_inst) {
return false;
}
if (!object_inst->type_id()) {
// We can only apply OpCopyObject to instructions that have types.
return false;
}
if (!context->get_decoration_mgr()
->GetDecorationsFor(message_.object(), true)
.empty()) {
// We do not copy objects that have decorations: if the copy is not
// decorated analogously, using the original object vs. its copy may not be
// equivalent.
// TODO(afd): it would be possible to make the copy but not add an id
// synonym.
if (!IsCopyable(context, object_inst)) {
return false;
}
@ -81,22 +70,11 @@ bool TransformationCopyObject::IsApplicable(
// The offset was inappropriate.
return false;
}
if (insert_before->PreviousNode() &&
(insert_before->PreviousNode()->opcode() == SpvOpLoopMerge ||
insert_before->PreviousNode()->opcode() == SpvOpSelectionMerge)) {
// We cannot insert a copy directly after a merge instruction.
return false;
}
if (insert_before->opcode() == SpvOpVariable) {
// We cannot insert a copy directly before a variable; variables in a
// function must be contiguous in the entry block.
return false;
}
// We cannot insert a copy directly before OpPhi, because OpPhi instructions
// need to be contiguous at the start of a block.
if (insert_before->opcode() == SpvOpPhi) {
if (!CanInsertCopyBefore(insert_before)) {
return false;
}
// |message_object| must be available at the point where we want to add the
// copy. It is available if it is at global scope (in which case it has no
// block), or if it dominates the point of insertion but is different from the
@ -154,5 +132,43 @@ protobufs::Transformation TransformationCopyObject::ToMessage() const {
return result;
}
bool TransformationCopyObject::IsCopyable(opt::IRContext* ir_context,
opt::Instruction* inst) {
if (!inst->HasResultId()) {
// We can only apply OpCopyObject to instructions that generate ids.
return false;
}
if (!inst->type_id()) {
// We can only apply OpCopyObject to instructions that have types.
return false;
}
// We do not copy objects that have decorations: if the copy is not
// decorated analogously, using the original object vs. its copy may not be
// equivalent.
// TODO(afd): it would be possible to make the copy but not add an id
// synonym.
return ir_context->get_decoration_mgr()
->GetDecorationsFor(inst->result_id(), true)
.empty();
}
bool TransformationCopyObject::CanInsertCopyBefore(
const opt::BasicBlock::iterator& instruction_in_block) {
if (instruction_in_block->PreviousNode() &&
(instruction_in_block->PreviousNode()->opcode() == SpvOpLoopMerge ||
instruction_in_block->PreviousNode()->opcode() == SpvOpSelectionMerge)) {
// We cannot insert a copy directly after a merge instruction.
return false;
}
if (instruction_in_block->opcode() == SpvOpVariable) {
// We cannot insert a copy directly before a variable; variables in a
// function must be contiguous in the entry block.
return false;
}
// We cannot insert a copy directly before OpPhi, because OpPhi instructions
// need to be contiguous at the start of a block.
return instruction_in_block->opcode() != SpvOpPhi;
}
} // namespace fuzz
} // namespace spvtools

Просмотреть файл

@ -28,8 +28,8 @@ class TransformationCopyObject : public Transformation {
explicit TransformationCopyObject(
const protobufs::TransformationCopyObject& message);
TransformationCopyObject(uint32_t fresh_id, uint32_t object,
uint32_t insert_after_id, uint32_t offset);
TransformationCopyObject(uint32_t object, uint32_t base_instruction_id,
uint32_t offset, uint32_t fresh_id);
// - |message_.fresh_id| must not be used by the module.
// - |message_.object| must be a result id that is a legitimate operand for
@ -58,6 +58,14 @@ class TransformationCopyObject : public Transformation {
protobufs::Transformation ToMessage() const override;
// Determines whether it is OK to make a copy of |inst|.
static bool IsCopyable(opt::IRContext* ir_context, opt::Instruction* inst);
// Determines whether it is OK to insert a copy instruction before the given
// instruction.
static bool CanInsertCopyBefore(
const opt::BasicBlock::iterator& instruction_in_block);
private:
protobufs::TransformationCopyObject message_;
};