SPIRV-Tools/source/fuzz/transformation_outline_function.cpp
Alastair Donaldson f278b467df
spirv-fuzz: Do not outline regions that end with a loop header (#3312)
The outliner would outline regions ending with a loop header, making
the block containing the call to the outlined function serve as the
loop header.  This, however, is incorrect in general, since the whole
outlined function -- rather than just the exit block for the region --
would end up getting called every time the loop would iterate.

This change restricts the outliner so that the last block in a region
cannot be a loop header.
2020-04-28 13:32:06 +01:00

968 lines
40 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/transformation_outline_function.h"
#include <set>
#include "source/fuzz/fuzzer_util.h"
namespace spvtools {
namespace fuzz {
namespace {
std::map<uint32_t, uint32_t> PairSequenceToMap(
const google::protobuf::RepeatedPtrField<protobufs::UInt32Pair>&
pair_sequence) {
std::map<uint32_t, uint32_t> result;
for (auto& pair : pair_sequence) {
result[pair.first()] = pair.second();
}
return result;
}
} // namespace
TransformationOutlineFunction::TransformationOutlineFunction(
const spvtools::fuzz::protobufs::TransformationOutlineFunction& message)
: message_(message) {}
TransformationOutlineFunction::TransformationOutlineFunction(
uint32_t entry_block, uint32_t exit_block,
uint32_t new_function_struct_return_type_id, uint32_t new_function_type_id,
uint32_t new_function_id, uint32_t new_function_region_entry_block,
uint32_t new_caller_result_id, uint32_t new_callee_result_id,
std::map<uint32_t, uint32_t>&& input_id_to_fresh_id,
std::map<uint32_t, uint32_t>&& output_id_to_fresh_id) {
message_.set_entry_block(entry_block);
message_.set_exit_block(exit_block);
message_.set_new_function_struct_return_type_id(
new_function_struct_return_type_id);
message_.set_new_function_type_id(new_function_type_id);
message_.set_new_function_id(new_function_id);
message_.set_new_function_region_entry_block(new_function_region_entry_block);
message_.set_new_caller_result_id(new_caller_result_id);
message_.set_new_callee_result_id(new_callee_result_id);
for (auto& entry : input_id_to_fresh_id) {
protobufs::UInt32Pair pair;
pair.set_first(entry.first);
pair.set_second(entry.second);
*message_.add_input_id_to_fresh_id() = pair;
}
for (auto& entry : output_id_to_fresh_id) {
protobufs::UInt32Pair pair;
pair.set_first(entry.first);
pair.set_second(entry.second);
*message_.add_output_id_to_fresh_id() = pair;
}
}
bool TransformationOutlineFunction::IsApplicable(
opt::IRContext* ir_context, const TransformationContext& /*unused*/) const {
std::set<uint32_t> ids_used_by_this_transformation;
// The various new ids used by the transformation must be fresh and distinct.
if (!CheckIdIsFreshAndNotUsedByThisTransformation(
message_.new_function_struct_return_type_id(), ir_context,
&ids_used_by_this_transformation)) {
return false;
}
if (!CheckIdIsFreshAndNotUsedByThisTransformation(
message_.new_function_type_id(), ir_context,
&ids_used_by_this_transformation)) {
return false;
}
if (!CheckIdIsFreshAndNotUsedByThisTransformation(
message_.new_function_id(), ir_context,
&ids_used_by_this_transformation)) {
return false;
}
if (!CheckIdIsFreshAndNotUsedByThisTransformation(
message_.new_function_region_entry_block(), ir_context,
&ids_used_by_this_transformation)) {
return false;
}
if (!CheckIdIsFreshAndNotUsedByThisTransformation(
message_.new_caller_result_id(), ir_context,
&ids_used_by_this_transformation)) {
return false;
}
if (!CheckIdIsFreshAndNotUsedByThisTransformation(
message_.new_callee_result_id(), ir_context,
&ids_used_by_this_transformation)) {
return false;
}
for (auto& pair : message_.input_id_to_fresh_id()) {
if (!CheckIdIsFreshAndNotUsedByThisTransformation(
pair.second(), ir_context, &ids_used_by_this_transformation)) {
return false;
}
}
for (auto& pair : message_.output_id_to_fresh_id()) {
if (!CheckIdIsFreshAndNotUsedByThisTransformation(
pair.second(), ir_context, &ids_used_by_this_transformation)) {
return false;
}
}
// The entry and exit block ids must indeed refer to blocks.
for (auto block_id : {message_.entry_block(), message_.exit_block()}) {
auto block_label = ir_context->get_def_use_mgr()->GetDef(block_id);
if (!block_label || block_label->opcode() != SpvOpLabel) {
return false;
}
}
auto entry_block = ir_context->cfg()->block(message_.entry_block());
auto exit_block = ir_context->cfg()->block(message_.exit_block());
// The entry block cannot start with OpVariable - this would mean that
// outlining would remove a variable from the function containing the region
// being outlined.
if (entry_block->begin()->opcode() == SpvOpVariable) {
return false;
}
// For simplicity, we do not allow the entry block to be a loop header.
if (entry_block->GetLoopMergeInst()) {
return false;
}
// For simplicity, we do not allow the exit block to be a merge block or
// continue target.
if (fuzzerutil::IsMergeOrContinue(ir_context, exit_block->id())) {
return false;
}
// The entry block cannot start with OpPhi. This is to keep the
// transformation logic simple. (Another transformation to split the OpPhis
// from a block could be applied to avoid this scenario.)
if (entry_block->begin()->opcode() == SpvOpPhi) {
return false;
}
// The block must be in the same function.
if (entry_block->GetParent() != exit_block->GetParent()) {
return false;
}
// The entry block must dominate the exit block.
auto dominator_analysis =
ir_context->GetDominatorAnalysis(entry_block->GetParent());
if (!dominator_analysis->Dominates(entry_block, exit_block)) {
return false;
}
// The exit block must post-dominate the entry block.
auto postdominator_analysis =
ir_context->GetPostDominatorAnalysis(entry_block->GetParent());
if (!postdominator_analysis->Dominates(exit_block, entry_block)) {
return false;
}
// Find all the blocks dominated by |message_.entry_block| and post-dominated
// by |message_.exit_block|.
auto region_set = GetRegionBlocks(
ir_context,
entry_block = ir_context->cfg()->block(message_.entry_block()),
exit_block = ir_context->cfg()->block(message_.exit_block()));
// Check whether |region_set| really is a single-entry single-exit region, and
// also check whether structured control flow constructs and their merge
// and continue constructs are either wholly in or wholly out of the region -
// e.g. avoid the situation where the region contains the head of a loop but
// not the loop's continue construct.
//
// This is achieved by going through every block in the function that contains
// the region.
for (auto& block : *entry_block->GetParent()) {
if (&block == exit_block) {
// It is OK (and typically expected) for the exit block of the region to
// have successors outside the region.
//
// It is also OK for the exit block to head a selection construct: the
// block containing the call to the outlined function will end up heading
// this construct if outlining takes place. However, it is not OK for
// the exit block to head a loop construct.
if (block.GetLoopMergeInst()) {
return false;
}
continue;
}
if (region_set.count(&block) != 0) {
// The block is in the region and is not the region's exit block. Let's
// see whether all of the block's successors are in the region. If they
// are not, the region is not single-entry single-exit.
bool all_successors_in_region = true;
block.WhileEachSuccessorLabel([&all_successors_in_region, ir_context,
&region_set](uint32_t successor) -> bool {
if (region_set.count(ir_context->cfg()->block(successor)) == 0) {
all_successors_in_region = false;
return false;
}
return true;
});
if (!all_successors_in_region) {
return false;
}
}
if (auto merge = block.GetMergeInst()) {
// The block is a loop or selection header -- the header and its
// associated merge block had better both be in the region or both be
// outside the region.
auto merge_block =
ir_context->cfg()->block(merge->GetSingleWordOperand(0));
if (region_set.count(&block) != region_set.count(merge_block)) {
return false;
}
}
if (auto loop_merge = block.GetLoopMergeInst()) {
// Similar to the above, but for the continue target of a loop.
auto continue_target =
ir_context->cfg()->block(loop_merge->GetSingleWordOperand(1));
if (continue_target != exit_block &&
region_set.count(&block) != region_set.count(continue_target)) {
return false;
}
}
}
// For each region input id, i.e. every id defined outside the region but
// used inside the region, ...
std::map<uint32_t, uint32_t> input_id_to_fresh_id_map =
PairSequenceToMap(message_.input_id_to_fresh_id());
for (auto id : GetRegionInputIds(ir_context, region_set, exit_block)) {
// There needs to be a corresponding fresh id to be used as a function
// parameter.
if (input_id_to_fresh_id_map.count(id) == 0) {
return false;
}
// Furthermore, if the input id has pointer type it must be an OpVariable
// or OpFunctionParameter.
auto input_id_inst = ir_context->get_def_use_mgr()->GetDef(id);
if (ir_context->get_def_use_mgr()
->GetDef(input_id_inst->type_id())
->opcode() == SpvOpTypePointer) {
switch (input_id_inst->opcode()) {
case SpvOpFunctionParameter:
case SpvOpVariable:
// These are OK.
break;
default:
// Anything else is not OK.
return false;
}
}
}
// For each region output id -- i.e. every id defined inside the region but
// used outside the region, ...
std::map<uint32_t, uint32_t> output_id_to_fresh_id_map =
PairSequenceToMap(message_.output_id_to_fresh_id());
for (auto id : GetRegionOutputIds(ir_context, region_set, exit_block)) {
if (
// ... there needs to be a corresponding fresh id that can hold the
// value for this id computed in the outlined function, and ...
output_id_to_fresh_id_map.count(id) == 0
// ... the output id must not have pointer type (to avoid creating a
// struct with pointer members to pass data out of the outlined
// function)
|| ir_context->get_def_use_mgr()
->GetDef(fuzzerutil::GetTypeId(ir_context, id))
->opcode() == SpvOpTypePointer) {
return false;
}
}
return true;
}
void TransformationOutlineFunction::Apply(
opt::IRContext* ir_context,
TransformationContext* transformation_context) const {
// The entry block for the region before outlining.
auto original_region_entry_block =
ir_context->cfg()->block(message_.entry_block());
// The exit block for the region before outlining.
auto original_region_exit_block =
ir_context->cfg()->block(message_.exit_block());
// The single-entry single-exit region defined by |message_.entry_block| and
// |message_.exit_block|.
std::set<opt::BasicBlock*> region_blocks = GetRegionBlocks(
ir_context, original_region_entry_block, original_region_exit_block);
// Input and output ids for the region being outlined.
std::vector<uint32_t> region_input_ids =
GetRegionInputIds(ir_context, region_blocks, original_region_exit_block);
std::vector<uint32_t> region_output_ids =
GetRegionOutputIds(ir_context, region_blocks, original_region_exit_block);
// Maps from input and output ids to fresh ids.
std::map<uint32_t, uint32_t> input_id_to_fresh_id_map =
PairSequenceToMap(message_.input_id_to_fresh_id());
std::map<uint32_t, uint32_t> output_id_to_fresh_id_map =
PairSequenceToMap(message_.output_id_to_fresh_id());
UpdateModuleIdBoundForFreshIds(ir_context, input_id_to_fresh_id_map,
output_id_to_fresh_id_map);
// Construct a map that associates each output id with its type id.
std::map<uint32_t, uint32_t> output_id_to_type_id;
for (uint32_t output_id : region_output_ids) {
output_id_to_type_id[output_id] =
ir_context->get_def_use_mgr()->GetDef(output_id)->type_id();
}
// The region will be collapsed to a single block that calls a function
// containing the outlined region. This block needs to end with whatever
// the exit block of the region ended with before outlining. We thus clone
// the terminator of the region's exit block, and the merge instruction for
// the block if there is one, so that we can append them to the end of the
// collapsed block later.
std::unique_ptr<opt::Instruction> cloned_exit_block_terminator =
std::unique_ptr<opt::Instruction>(
original_region_exit_block->terminator()->Clone(ir_context));
std::unique_ptr<opt::Instruction> cloned_exit_block_merge =
original_region_exit_block->GetMergeInst()
? std::unique_ptr<opt::Instruction>(
original_region_exit_block->GetMergeInst()->Clone(ir_context))
: nullptr;
// Make a function prototype for the outlined function, which involves
// figuring out its required type.
std::unique_ptr<opt::Function> outlined_function = PrepareFunctionPrototype(
region_input_ids, region_output_ids, input_id_to_fresh_id_map, ir_context,
transformation_context);
// If the original function was livesafe, the new function should also be
// livesafe.
if (transformation_context->GetFactManager()->FunctionIsLivesafe(
original_region_entry_block->GetParent()->result_id())) {
transformation_context->GetFactManager()->AddFactFunctionIsLivesafe(
message_.new_function_id());
}
// Adapt the region to be outlined so that its input ids are replaced with the
// ids of the outlined function's input parameters, and so that output ids
// are similarly remapped.
RemapInputAndOutputIdsInRegion(
ir_context, *original_region_exit_block, region_blocks, region_input_ids,
region_output_ids, input_id_to_fresh_id_map, output_id_to_fresh_id_map);
// Fill out the body of the outlined function according to the region that is
// being outlined.
PopulateOutlinedFunction(
*original_region_entry_block, *original_region_exit_block, region_blocks,
region_output_ids, output_id_to_fresh_id_map, ir_context,
outlined_function.get(), transformation_context);
// Collapse the region that has been outlined into a function down to a single
// block that calls said function.
ShrinkOriginalRegion(
ir_context, region_blocks, region_input_ids, region_output_ids,
output_id_to_type_id, outlined_function->type_id(),
std::move(cloned_exit_block_merge),
std::move(cloned_exit_block_terminator), original_region_entry_block);
// Add the outlined function to the module.
ir_context->module()->AddFunction(std::move(outlined_function));
// Major surgery has been conducted on the module, so invalidate all analyses.
ir_context->InvalidateAnalysesExceptFor(
opt::IRContext::Analysis::kAnalysisNone);
}
protobufs::Transformation TransformationOutlineFunction::ToMessage() const {
protobufs::Transformation result;
*result.mutable_outline_function() = message_;
return result;
}
std::vector<uint32_t> TransformationOutlineFunction::GetRegionInputIds(
opt::IRContext* ir_context, const std::set<opt::BasicBlock*>& region_set,
opt::BasicBlock* region_exit_block) {
std::vector<uint32_t> result;
auto enclosing_function = region_exit_block->GetParent();
// Consider each parameter of the function containing the region.
enclosing_function->ForEachParam(
[ir_context, &region_set, &result](opt::Instruction* function_parameter) {
// Consider every use of the parameter.
ir_context->get_def_use_mgr()->WhileEachUse(
function_parameter,
[ir_context, function_parameter, &region_set, &result](
opt::Instruction* use, uint32_t /*unused*/) {
// Get the block, if any, in which the parameter is used.
auto use_block = ir_context->get_instr_block(use);
// If the use is in a block that lies within the region, the
// parameter is an input id for the region.
if (use_block && region_set.count(use_block) != 0) {
result.push_back(function_parameter->result_id());
return false;
}
return true;
});
});
// Consider all definitions in the function that might turn out to be input
// ids.
for (auto& block : *enclosing_function) {
std::vector<opt::Instruction*> candidate_input_ids_for_block;
if (region_set.count(&block) == 0) {
// All instructions in blocks outside the region are candidate's for
// generating input ids.
for (auto& inst : block) {
candidate_input_ids_for_block.push_back(&inst);
}
} else {
// Blocks in the region cannot generate input ids.
continue;
}
// Consider each candidate input id to check whether it is used in the
// region.
for (auto& inst : candidate_input_ids_for_block) {
ir_context->get_def_use_mgr()->WhileEachUse(
inst,
[ir_context, &inst, region_exit_block, &region_set, &result](
opt::Instruction* use, uint32_t /*unused*/) -> bool {
// Find the block in which this id use occurs, recording the id as
// an input id if the block is outside the region, with some
// exceptions detailed below.
auto use_block = ir_context->get_instr_block(use);
if (!use_block) {
// There might be no containing block, e.g. if the use is in a
// decoration.
return true;
}
if (region_set.count(use_block) == 0) {
// The use is not in the region: this does not make it an input
// id.
return true;
}
if (use_block == region_exit_block && use->IsBlockTerminator()) {
// We do not regard uses in the exit block terminator as input
// ids, as this terminator does not get outlined.
return true;
}
result.push_back(inst->result_id());
return false;
});
}
}
return result;
}
std::vector<uint32_t> TransformationOutlineFunction::GetRegionOutputIds(
opt::IRContext* ir_context, const std::set<opt::BasicBlock*>& region_set,
opt::BasicBlock* region_exit_block) {
std::vector<uint32_t> result;
// Consider each block in the function containing the region.
for (auto& block : *region_exit_block->GetParent()) {
if (region_set.count(&block) == 0) {
// Skip blocks that are not in the region.
continue;
}
// Consider each use of each instruction defined in the block.
for (auto& inst : block) {
ir_context->get_def_use_mgr()->WhileEachUse(
&inst,
[&region_set, ir_context, &inst, region_exit_block, &result](
opt::Instruction* use, uint32_t /*unused*/) -> bool {
// Find the block in which this id use occurs, recording the id as
// an output id if the block is outside the region, with some
// exceptions detailed below.
auto use_block = ir_context->get_instr_block(use);
if (!use_block) {
// There might be no containing block, e.g. if the use is in a
// decoration.
return true;
}
if (region_set.count(use_block) != 0) {
// The use is in the region.
if (use_block != region_exit_block || !use->IsBlockTerminator()) {
// Furthermore, the use is not in the terminator of the region's
// exit block.
return true;
}
}
result.push_back(inst.result_id());
return false;
});
}
}
return result;
}
std::set<opt::BasicBlock*> TransformationOutlineFunction::GetRegionBlocks(
opt::IRContext* ir_context, opt::BasicBlock* entry_block,
opt::BasicBlock* exit_block) {
auto enclosing_function = entry_block->GetParent();
auto dominator_analysis =
ir_context->GetDominatorAnalysis(enclosing_function);
auto postdominator_analysis =
ir_context->GetPostDominatorAnalysis(enclosing_function);
std::set<opt::BasicBlock*> result;
for (auto& block : *enclosing_function) {
if (dominator_analysis->Dominates(entry_block, &block) &&
postdominator_analysis->Dominates(exit_block, &block)) {
result.insert(&block);
}
}
return result;
}
std::unique_ptr<opt::Function>
TransformationOutlineFunction::PrepareFunctionPrototype(
const std::vector<uint32_t>& region_input_ids,
const std::vector<uint32_t>& region_output_ids,
const std::map<uint32_t, uint32_t>& input_id_to_fresh_id_map,
opt::IRContext* ir_context,
TransformationContext* transformation_context) const {
uint32_t return_type_id = 0;
uint32_t function_type_id = 0;
// First, try to find an existing function type that is suitable. This is
// only possible if the region generates no output ids; if it generates output
// ids we are going to make a new struct for those, and since that struct does
// not exist there cannot already be a function type with this struct as its
// return type.
if (region_output_ids.empty()) {
std::vector<uint32_t> return_and_parameter_types;
opt::analysis::Void void_type;
return_type_id = ir_context->get_type_mgr()->GetId(&void_type);
return_and_parameter_types.push_back(return_type_id);
for (auto id : region_input_ids) {
return_and_parameter_types.push_back(
ir_context->get_def_use_mgr()->GetDef(id)->type_id());
}
function_type_id =
fuzzerutil::FindFunctionType(ir_context, return_and_parameter_types);
}
// If no existing function type was found, we need to create one.
if (function_type_id == 0) {
assert(
((return_type_id == 0) == !region_output_ids.empty()) &&
"We should only have set the return type if there are no output ids.");
// If the region generates output ids, we need to make a struct with one
// field per output id.
if (!region_output_ids.empty()) {
opt::Instruction::OperandList struct_member_types;
for (uint32_t output_id : region_output_ids) {
auto output_id_type =
ir_context->get_def_use_mgr()->GetDef(output_id)->type_id();
struct_member_types.push_back({SPV_OPERAND_TYPE_ID, {output_id_type}});
}
// Add a new struct type to the module.
ir_context->module()->AddType(MakeUnique<opt::Instruction>(
ir_context, SpvOpTypeStruct, 0,
message_.new_function_struct_return_type_id(),
std::move(struct_member_types)));
// The return type for the function is the newly-created struct.
return_type_id = message_.new_function_struct_return_type_id();
}
assert(
return_type_id != 0 &&
"We should either have a void return type, or have created a struct.");
// The region's input ids dictate the parameter types to the function.
opt::Instruction::OperandList function_type_operands;
function_type_operands.push_back({SPV_OPERAND_TYPE_ID, {return_type_id}});
for (auto id : region_input_ids) {
function_type_operands.push_back(
{SPV_OPERAND_TYPE_ID,
{ir_context->get_def_use_mgr()->GetDef(id)->type_id()}});
}
// Add a new function type to the module, and record that this is the type
// id for the new function.
ir_context->module()->AddType(MakeUnique<opt::Instruction>(
ir_context, SpvOpTypeFunction, 0, message_.new_function_type_id(),
function_type_operands));
function_type_id = message_.new_function_type_id();
}
// Create a new function with |message_.new_function_id| as the function id,
// and the return type and function type prepared above.
std::unique_ptr<opt::Function> outlined_function =
MakeUnique<opt::Function>(MakeUnique<opt::Instruction>(
ir_context, SpvOpFunction, return_type_id, message_.new_function_id(),
opt::Instruction::OperandList(
{{spv_operand_type_t ::SPV_OPERAND_TYPE_LITERAL_INTEGER,
{SpvFunctionControlMaskNone}},
{spv_operand_type_t::SPV_OPERAND_TYPE_ID,
{function_type_id}}})));
// Add one parameter to the function for each input id, using the fresh ids
// provided in |input_id_to_fresh_id_map|.
for (auto id : region_input_ids) {
outlined_function->AddParameter(MakeUnique<opt::Instruction>(
ir_context, SpvOpFunctionParameter,
ir_context->get_def_use_mgr()->GetDef(id)->type_id(),
input_id_to_fresh_id_map.at(id), opt::Instruction::OperandList()));
// If the input id is an irrelevant-valued variable, the same should be true
// of the corresponding parameter.
if (transformation_context->GetFactManager()->PointeeValueIsIrrelevant(
id)) {
transformation_context->GetFactManager()
->AddFactValueOfPointeeIsIrrelevant(input_id_to_fresh_id_map.at(id));
}
}
return outlined_function;
}
void TransformationOutlineFunction::UpdateModuleIdBoundForFreshIds(
opt::IRContext* ir_context,
const std::map<uint32_t, uint32_t>& input_id_to_fresh_id_map,
const std::map<uint32_t, uint32_t>& output_id_to_fresh_id_map) const {
// Enlarge the module's id bound as needed to accommodate the various fresh
// ids associated with the transformation.
fuzzerutil::UpdateModuleIdBound(
ir_context, message_.new_function_struct_return_type_id());
fuzzerutil::UpdateModuleIdBound(ir_context, message_.new_function_type_id());
fuzzerutil::UpdateModuleIdBound(ir_context, message_.new_function_id());
fuzzerutil::UpdateModuleIdBound(ir_context,
message_.new_function_region_entry_block());
fuzzerutil::UpdateModuleIdBound(ir_context, message_.new_caller_result_id());
fuzzerutil::UpdateModuleIdBound(ir_context, message_.new_callee_result_id());
for (auto& entry : input_id_to_fresh_id_map) {
fuzzerutil::UpdateModuleIdBound(ir_context, entry.second);
}
for (auto& entry : output_id_to_fresh_id_map) {
fuzzerutil::UpdateModuleIdBound(ir_context, entry.second);
}
}
void TransformationOutlineFunction::RemapInputAndOutputIdsInRegion(
opt::IRContext* ir_context,
const opt::BasicBlock& original_region_exit_block,
const std::set<opt::BasicBlock*>& region_blocks,
const std::vector<uint32_t>& region_input_ids,
const std::vector<uint32_t>& region_output_ids,
const std::map<uint32_t, uint32_t>& input_id_to_fresh_id_map,
const std::map<uint32_t, uint32_t>& output_id_to_fresh_id_map) const {
// Change all uses of input ids inside the region to the corresponding fresh
// ids that will ultimately be parameters of the outlined function.
// This is done by considering each region input id in turn.
for (uint32_t id : region_input_ids) {
// We then consider each use of the input id.
ir_context->get_def_use_mgr()->ForEachUse(
id, [ir_context, id, &input_id_to_fresh_id_map, region_blocks](
opt::Instruction* use, uint32_t operand_index) {
// Find the block in which this use of the input id occurs.
opt::BasicBlock* use_block = ir_context->get_instr_block(use);
// We want to rewrite the use id if its block occurs in the outlined
// region.
if (region_blocks.count(use_block) != 0) {
// Rewrite this use of the input id.
use->SetOperand(operand_index, {input_id_to_fresh_id_map.at(id)});
}
});
}
// Change each definition of a region output id to define the corresponding
// fresh ids that will store intermediate value for the output ids. Also
// change all uses of the output id located in the outlined region.
// This is done by considering each region output id in turn.
for (uint32_t id : region_output_ids) {
// First consider each use of the output id and update the relevant uses.
ir_context->get_def_use_mgr()->ForEachUse(
id, [ir_context, &original_region_exit_block, id,
&output_id_to_fresh_id_map,
region_blocks](opt::Instruction* use, uint32_t operand_index) {
// Find the block in which this use of the output id occurs.
auto use_block = ir_context->get_instr_block(use);
// We want to rewrite the use id if its block occurs in the outlined
// region, with one exception: the terminator of the exit block of
// the region is going to remain in the original function, so if the
// use appears in such a terminator instruction we leave it alone.
if (
// The block is in the region ...
region_blocks.count(use_block) != 0 &&
// ... and the use is not in the terminator instruction of the
// region's exit block.
!(use_block == &original_region_exit_block &&
use->IsBlockTerminator())) {
// Rewrite this use of the output id.
use->SetOperand(operand_index, {output_id_to_fresh_id_map.at(id)});
}
});
// Now change the instruction that defines the output id so that it instead
// defines the corresponding fresh id. We do this after changing all the
// uses so that the definition of the original id is still registered when
// we analyse its uses.
ir_context->get_def_use_mgr()->GetDef(id)->SetResultId(
output_id_to_fresh_id_map.at(id));
}
}
void TransformationOutlineFunction::PopulateOutlinedFunction(
const opt::BasicBlock& original_region_entry_block,
const opt::BasicBlock& original_region_exit_block,
const std::set<opt::BasicBlock*>& region_blocks,
const std::vector<uint32_t>& region_output_ids,
const std::map<uint32_t, uint32_t>& output_id_to_fresh_id_map,
opt::IRContext* ir_context, opt::Function* outlined_function,
TransformationContext* transformation_context) const {
// When we create the exit block for the outlined region, we use this pointer
// to track of it so that we can manipulate it later.
opt::BasicBlock* outlined_region_exit_block = nullptr;
// The region entry block in the new function is identical to the entry block
// of the region being outlined, except that it has
// |message_.new_function_region_entry_block| as its id.
std::unique_ptr<opt::BasicBlock> outlined_region_entry_block =
MakeUnique<opt::BasicBlock>(MakeUnique<opt::Instruction>(
ir_context, SpvOpLabel, 0, message_.new_function_region_entry_block(),
opt::Instruction::OperandList()));
outlined_region_entry_block->SetParent(outlined_function);
// If the original region's entry block was dead, the outlined region's entry
// block is also dead.
if (transformation_context->GetFactManager()->BlockIsDead(
original_region_entry_block.id())) {
transformation_context->GetFactManager()->AddFactBlockIsDead(
outlined_region_entry_block->id());
}
if (&original_region_entry_block == &original_region_exit_block) {
outlined_region_exit_block = outlined_region_entry_block.get();
}
for (auto& inst : original_region_entry_block) {
outlined_region_entry_block->AddInstruction(
std::unique_ptr<opt::Instruction>(inst.Clone(ir_context)));
}
outlined_function->AddBasicBlock(std::move(outlined_region_entry_block));
// We now go through the single-entry single-exit region defined by the entry
// and exit blocks, adding clones of all blocks to the new function.
// Consider every block in the enclosing function.
auto enclosing_function = original_region_entry_block.GetParent();
for (auto block_it = enclosing_function->begin();
block_it != enclosing_function->end();) {
// Skip the region's entry block - we already dealt with it above.
if (region_blocks.count(&*block_it) == 0 ||
&*block_it == &original_region_entry_block) {
++block_it;
continue;
}
// Clone the block so that it can be added to the new function.
auto cloned_block =
std::unique_ptr<opt::BasicBlock>(block_it->Clone(ir_context));
// If this is the region's exit block, then the cloned block is the outlined
// region's exit block.
if (&*block_it == &original_region_exit_block) {
assert(outlined_region_exit_block == nullptr &&
"We should not yet have encountered the exit block.");
outlined_region_exit_block = cloned_block.get();
}
cloned_block->SetParent(outlined_function);
// Redirect any OpPhi operands whose predecessors are the original region
// entry block to become the new function entry block.
cloned_block->ForEachPhiInst([this](opt::Instruction* phi_inst) {
for (uint32_t predecessor_index = 1;
predecessor_index < phi_inst->NumInOperands();
predecessor_index += 2) {
if (phi_inst->GetSingleWordInOperand(predecessor_index) ==
message_.entry_block()) {
phi_inst->SetInOperand(predecessor_index,
{message_.new_function_region_entry_block()});
}
}
});
outlined_function->AddBasicBlock(std::move(cloned_block));
block_it = block_it.Erase();
}
assert(outlined_region_exit_block != nullptr &&
"We should have encountered the region's exit block when iterating "
"through the function");
// We now need to adapt the exit block for the region - in the new function -
// so that it ends with a return.
// We first eliminate the merge instruction (if any) and the terminator for
// the cloned exit block.
for (auto inst_it = outlined_region_exit_block->begin();
inst_it != outlined_region_exit_block->end();) {
if (inst_it->opcode() == SpvOpLoopMerge ||
inst_it->opcode() == SpvOpSelectionMerge) {
inst_it = inst_it.Erase();
} else if (inst_it->IsBlockTerminator()) {
inst_it = inst_it.Erase();
} else {
++inst_it;
}
}
// We now add either OpReturn or OpReturnValue as the cloned exit block's
// terminator.
if (region_output_ids.empty()) {
// The case where there are no region output ids is simple: we just add
// OpReturn.
outlined_region_exit_block->AddInstruction(MakeUnique<opt::Instruction>(
ir_context, SpvOpReturn, 0, 0, opt::Instruction::OperandList()));
} else {
// In the case where there are output ids, we add an OpCompositeConstruct
// instruction to pack all the output values into a struct, and then an
// OpReturnValue instruction to return this struct.
opt::Instruction::OperandList struct_member_operands;
for (uint32_t id : region_output_ids) {
struct_member_operands.push_back(
{SPV_OPERAND_TYPE_ID, {output_id_to_fresh_id_map.at(id)}});
}
outlined_region_exit_block->AddInstruction(MakeUnique<opt::Instruction>(
ir_context, SpvOpCompositeConstruct,
message_.new_function_struct_return_type_id(),
message_.new_callee_result_id(), struct_member_operands));
outlined_region_exit_block->AddInstruction(MakeUnique<opt::Instruction>(
ir_context, SpvOpReturnValue, 0, 0,
opt::Instruction::OperandList(
{{SPV_OPERAND_TYPE_ID, {message_.new_callee_result_id()}}})));
}
outlined_function->SetFunctionEnd(MakeUnique<opt::Instruction>(
ir_context, SpvOpFunctionEnd, 0, 0, opt::Instruction::OperandList()));
}
void TransformationOutlineFunction::ShrinkOriginalRegion(
opt::IRContext* ir_context, std::set<opt::BasicBlock*>& region_blocks,
const std::vector<uint32_t>& region_input_ids,
const std::vector<uint32_t>& region_output_ids,
const std::map<uint32_t, uint32_t>& output_id_to_type_id,
uint32_t return_type_id,
std::unique_ptr<opt::Instruction> cloned_exit_block_merge,
std::unique_ptr<opt::Instruction> cloned_exit_block_terminator,
opt::BasicBlock* original_region_entry_block) const {
// Erase all blocks from the original function that are in the outlined
// region, except for the region's entry block.
//
// In the process, identify all references to the exit block of the region,
// as merge blocks, continue targets, or OpPhi predecessors, and rewrite them
// to refer to the region entry block (the single block to which we are
// shrinking the region).
auto enclosing_function = original_region_entry_block->GetParent();
for (auto block_it = enclosing_function->begin();
block_it != enclosing_function->end();) {
if (&*block_it == original_region_entry_block) {
++block_it;
} else if (region_blocks.count(&*block_it) == 0) {
// The block is not in the region. Check whether it has the last block
// of the region as an OpPhi predecessor, and if so change the
// predecessor to be the first block of the region (i.e. the block
// containing the call to what was outlined).
assert(block_it->MergeBlockIdIfAny() != message_.exit_block() &&
"Outlined region must not end with a merge block");
assert(block_it->ContinueBlockIdIfAny() != message_.exit_block() &&
"Outlined region must not end with a continue target");
block_it->ForEachPhiInst([this](opt::Instruction* phi_inst) {
for (uint32_t predecessor_index = 1;
predecessor_index < phi_inst->NumInOperands();
predecessor_index += 2) {
if (phi_inst->GetSingleWordInOperand(predecessor_index) ==
message_.exit_block()) {
phi_inst->SetInOperand(predecessor_index, {message_.entry_block()});
}
}
});
++block_it;
} else {
// The block is in the region and is not the region's entry block: kill
// it.
block_it = block_it.Erase();
}
}
// Now erase all instructions from the region's entry block, as they have
// been outlined.
for (auto inst_it = original_region_entry_block->begin();
inst_it != original_region_entry_block->end();) {
inst_it = inst_it.Erase();
}
// Now we add a call to the outlined function to the region's entry block.
opt::Instruction::OperandList function_call_operands;
function_call_operands.push_back(
{SPV_OPERAND_TYPE_ID, {message_.new_function_id()}});
// The function parameters are the region input ids.
for (auto input_id : region_input_ids) {
function_call_operands.push_back({SPV_OPERAND_TYPE_ID, {input_id}});
}
original_region_entry_block->AddInstruction(MakeUnique<opt::Instruction>(
ir_context, SpvOpFunctionCall, return_type_id,
message_.new_caller_result_id(), function_call_operands));
// If there are output ids, the function call will return a struct. For each
// output id, we add an extract operation to pull the appropriate struct
// member out into an output id.
for (uint32_t index = 0; index < region_output_ids.size(); ++index) {
uint32_t output_id = region_output_ids[index];
original_region_entry_block->AddInstruction(MakeUnique<opt::Instruction>(
ir_context, SpvOpCompositeExtract, output_id_to_type_id.at(output_id),
output_id,
opt::Instruction::OperandList(
{{SPV_OPERAND_TYPE_ID, {message_.new_caller_result_id()}},
{SPV_OPERAND_TYPE_LITERAL_INTEGER, {index}}})));
}
// Finally, we terminate the block with the merge instruction (if any) that
// used to belong to the region's exit block, and the terminator that used
// to belong to the region's exit block.
if (cloned_exit_block_merge != nullptr) {
original_region_entry_block->AddInstruction(
std::move(cloned_exit_block_merge));
}
original_region_entry_block->AddInstruction(
std::move(cloned_exit_block_terminator));
}
} // namespace fuzz
} // namespace spvtools