SPIRV-Tools/source/opt/loop_peeling.cpp

1084 lines
37 KiB
C++
Raw Normal View History

// Copyright (c) 2018 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 <algorithm>
#include <functional>
#include <memory>
#include <unordered_map>
#include <unordered_set>
#include <vector>
#include "source/opt/ir_builder.h"
#include "source/opt/ir_context.h"
#include "source/opt/loop_descriptor.h"
#include "source/opt/loop_peeling.h"
#include "source/opt/loop_utils.h"
#include "source/opt/scalar_analysis.h"
#include "source/opt/scalar_analysis_nodes.h"
namespace spvtools {
namespace opt {
size_t LoopPeelingPass::code_grow_threshold_ = 1000;
void LoopPeeling::DuplicateAndConnectLoop(
LoopUtils::LoopCloningResult* clone_results) {
CFG& cfg = *context_->cfg();
analysis::DefUseManager* def_use_mgr = context_->get_def_use_mgr();
assert(CanPeelLoop() && "Cannot peel loop!");
std::vector<BasicBlock*> ordered_loop_blocks;
BasicBlock* pre_header = loop_->GetOrCreatePreHeaderBlock();
loop_->ComputeLoopStructuredOrder(&ordered_loop_blocks);
cloned_loop_ = loop_utils_.CloneLoop(clone_results, ordered_loop_blocks);
// Add the basic block to the function.
Function::iterator it =
loop_utils_.GetFunction()->FindBlock(pre_header->id());
assert(it != loop_utils_.GetFunction()->end() &&
"Pre-header not found in the function.");
loop_utils_.GetFunction()->AddBasicBlocks(
clone_results->cloned_bb_.begin(), clone_results->cloned_bb_.end(), ++it);
// Make the |loop_|'s preheader the |cloned_loop_| one.
BasicBlock* cloned_header = cloned_loop_->GetHeaderBlock();
pre_header->ForEachSuccessorLabel(
[cloned_header](uint32_t* succ) { *succ = cloned_header->id(); });
// Update cfg.
cfg.RemoveEdge(pre_header->id(), loop_->GetHeaderBlock()->id());
cloned_loop_->SetPreHeaderBlock(pre_header);
loop_->SetPreHeaderBlock(nullptr);
// When cloning the loop, we didn't cloned the merge block, so currently
// |cloned_loop_| shares the same block as |loop_|.
// We mutate all branches from |cloned_loop_| block to |loop_|'s merge into a
// branch to |loop_|'s header (so header will also be the merge of
// |cloned_loop_|).
uint32_t cloned_loop_exit = 0;
for (uint32_t pred_id : cfg.preds(loop_->GetMergeBlock()->id())) {
if (loop_->IsInsideLoop(pred_id)) continue;
BasicBlock* bb = cfg.block(pred_id);
assert(cloned_loop_exit == 0 && "The loop has multiple exits.");
cloned_loop_exit = bb->id();
bb->ForEachSuccessorLabel([this](uint32_t* succ) {
if (*succ == loop_->GetMergeBlock()->id())
*succ = loop_->GetHeaderBlock()->id();
});
}
// Update cfg.
cfg.RemoveNonExistingEdges(loop_->GetMergeBlock()->id());
cfg.AddEdge(cloned_loop_exit, loop_->GetHeaderBlock()->id());
// Patch the phi of the original loop header:
// - Set the loop entry branch to come from the cloned loop exit block;
// - Set the initial value of the phi using the corresponding cloned loop
// exit values.
//
// We patch the iterating value initializers of the original loop using the
// corresponding cloned loop exit values. Connects the cloned loop iterating
// values to the original loop. This make sure that the initial value of the
// second loop starts with the last value of the first loop.
//
// For example, loops like:
//
// int z = 0;
// for (int i = 0; i++ < M; i += cst1) {
// if (cond)
// z += cst2;
// }
//
// Will become:
//
// int z = 0;
// int i = 0;
// for (; i++ < M; i += cst1) {
// if (cond)
// z += cst2;
// }
// for (; i++ < M; i += cst1) {
// if (cond)
// z += cst2;
// }
loop_->GetHeaderBlock()->ForEachPhiInst([cloned_loop_exit, def_use_mgr,
clone_results,
this](Instruction* phi) {
for (uint32_t i = 0; i < phi->NumInOperands(); i += 2) {
if (!loop_->IsInsideLoop(phi->GetSingleWordInOperand(i + 1))) {
phi->SetInOperand(i,
{clone_results->value_map_.at(
exit_value_.at(phi->result_id())->result_id())});
phi->SetInOperand(i + 1, {cloned_loop_exit});
def_use_mgr->AnalyzeInstUse(phi);
return;
}
}
});
// Force the creation of a new preheader for the original loop and set it as
// the merge block for the cloned loop.
cloned_loop_->SetMergeBlock(loop_->GetOrCreatePreHeaderBlock());
}
void LoopPeeling::InsertCanonicalInductionVariable(
LoopUtils::LoopCloningResult* clone_results) {
if (original_loop_canonical_induction_variable_) {
canonical_induction_variable_ =
context_->get_def_use_mgr()->GetDef(clone_results->value_map_.at(
original_loop_canonical_induction_variable_->result_id()));
return;
}
BasicBlock::iterator insert_point = GetClonedLoop()->GetLatchBlock()->tail();
if (GetClonedLoop()->GetLatchBlock()->GetMergeInst()) {
--insert_point;
}
InstructionBuilder builder(
context_, &*insert_point,
IRContext::kAnalysisDefUse | IRContext::kAnalysisInstrToBlockMapping);
Instruction* uint_1_cst =
builder.Add32BitConstantInteger<uint32_t>(1, int_type_->IsSigned());
// Create the increment.
// Note that we do "1 + 1" here, one of the operand should the phi
// value but we don't have it yet. The operand will be set latter.
Instruction* iv_inc = builder.AddIAdd(
uint_1_cst->type_id(), uint_1_cst->result_id(), uint_1_cst->result_id());
builder.SetInsertPoint(&*GetClonedLoop()->GetHeaderBlock()->begin());
canonical_induction_variable_ = builder.AddPhi(
uint_1_cst->type_id(),
{builder.Add32BitConstantInteger<uint32_t>(0, int_type_->IsSigned())
->result_id(),
GetClonedLoop()->GetPreHeaderBlock()->id(), iv_inc->result_id(),
GetClonedLoop()->GetLatchBlock()->id()});
// Connect everything.
iv_inc->SetInOperand(0, {canonical_induction_variable_->result_id()});
// Update def/use manager.
context_->get_def_use_mgr()->AnalyzeInstUse(iv_inc);
// If do-while form, use the incremented value.
if (do_while_form_) {
canonical_induction_variable_ = iv_inc;
}
}
void LoopPeeling::GetIteratorUpdateOperations(
const Loop* loop, Instruction* iterator,
std::unordered_set<Instruction*>* operations) {
analysis::DefUseManager* def_use_mgr = context_->get_def_use_mgr();
operations->insert(iterator);
iterator->ForEachInId([def_use_mgr, loop, operations, this](uint32_t* id) {
Instruction* insn = def_use_mgr->GetDef(*id);
if (insn->opcode() == SpvOpLabel) {
return;
}
if (operations->count(insn)) {
return;
}
if (!loop->IsInsideLoop(insn)) {
return;
}
GetIteratorUpdateOperations(loop, insn, operations);
});
}
// Gather the set of blocks for all the path from |entry| to |root|.
static void GetBlocksInPath(uint32_t block, uint32_t entry,
std::unordered_set<uint32_t>* blocks_in_path,
const CFG& cfg) {
for (uint32_t pid : cfg.preds(block)) {
if (blocks_in_path->insert(pid).second) {
if (pid != entry) {
GetBlocksInPath(pid, entry, blocks_in_path, cfg);
}
}
}
}
bool LoopPeeling::IsConditionCheckSideEffectFree() const {
CFG& cfg = *context_->cfg();
// The "do-while" form does not cause issues, the algorithm takes into account
// the first iteration.
if (!do_while_form_) {
uint32_t condition_block_id = cfg.preds(loop_->GetMergeBlock()->id())[0];
std::unordered_set<uint32_t> blocks_in_path;
blocks_in_path.insert(condition_block_id);
GetBlocksInPath(condition_block_id, loop_->GetHeaderBlock()->id(),
&blocks_in_path, cfg);
for (uint32_t bb_id : blocks_in_path) {
BasicBlock* bb = cfg.block(bb_id);
if (!bb->WhileEachInst([this](Instruction* insn) {
if (insn->IsBranch()) return true;
switch (insn->opcode()) {
case SpvOpLabel:
case SpvOpSelectionMerge:
case SpvOpLoopMerge:
return true;
default:
break;
}
return context_->IsCombinatorInstruction(insn);
})) {
return false;
}
}
}
return true;
}
void LoopPeeling::GetIteratingExitValues() {
CFG& cfg = *context_->cfg();
loop_->GetHeaderBlock()->ForEachPhiInst(
[this](Instruction* phi) { exit_value_[phi->result_id()] = nullptr; });
if (!loop_->GetMergeBlock()) {
return;
}
if (cfg.preds(loop_->GetMergeBlock()->id()).size() != 1) {
return;
}
analysis::DefUseManager* def_use_mgr = context_->get_def_use_mgr();
uint32_t condition_block_id = cfg.preds(loop_->GetMergeBlock()->id())[0];
auto& header_pred = cfg.preds(loop_->GetHeaderBlock()->id());
do_while_form_ = std::find(header_pred.begin(), header_pred.end(),
condition_block_id) != header_pred.end();
if (do_while_form_) {
loop_->GetHeaderBlock()->ForEachPhiInst(
[condition_block_id, def_use_mgr, this](Instruction* phi) {
std::unordered_set<Instruction*> operations;
for (uint32_t i = 0; i < phi->NumInOperands(); i += 2) {
if (condition_block_id == phi->GetSingleWordInOperand(i + 1)) {
exit_value_[phi->result_id()] =
def_use_mgr->GetDef(phi->GetSingleWordInOperand(i));
}
}
});
} else {
DominatorTree* dom_tree =
&context_->GetDominatorAnalysis(loop_utils_.GetFunction())
->GetDomTree();
BasicBlock* condition_block = cfg.block(condition_block_id);
loop_->GetHeaderBlock()->ForEachPhiInst(
[dom_tree, condition_block, this](Instruction* phi) {
std::unordered_set<Instruction*> operations;
// Not the back-edge value, check if the phi instruction is the only
// possible candidate.
GetIteratorUpdateOperations(loop_, phi, &operations);
for (Instruction* insn : operations) {
if (insn == phi) {
continue;
}
if (dom_tree->Dominates(context_->get_instr_block(insn),
condition_block)) {
return;
}
}
exit_value_[phi->result_id()] = phi;
});
}
}
void LoopPeeling::FixExitCondition(
const std::function<uint32_t(Instruction*)>& condition_builder) {
CFG& cfg = *context_->cfg();
uint32_t condition_block_id = 0;
for (uint32_t id : cfg.preds(GetClonedLoop()->GetMergeBlock()->id())) {
if (GetClonedLoop()->IsInsideLoop(id)) {
condition_block_id = id;
break;
}
}
assert(condition_block_id != 0 && "2nd loop in improperly connected");
BasicBlock* condition_block = cfg.block(condition_block_id);
Instruction* exit_condition = condition_block->terminator();
assert(exit_condition->opcode() == SpvOpBranchConditional);
BasicBlock::iterator insert_point = condition_block->tail();
if (condition_block->GetMergeInst()) {
--insert_point;
}
exit_condition->SetInOperand(0, {condition_builder(&*insert_point)});
uint32_t to_continue_block_idx =
GetClonedLoop()->IsInsideLoop(exit_condition->GetSingleWordInOperand(1))
? 1
: 2;
exit_condition->SetInOperand(
1, {exit_condition->GetSingleWordInOperand(to_continue_block_idx)});
exit_condition->SetInOperand(2, {GetClonedLoop()->GetMergeBlock()->id()});
// Update def/use manager.
context_->get_def_use_mgr()->AnalyzeInstUse(exit_condition);
}
BasicBlock* LoopPeeling::CreateBlockBefore(BasicBlock* bb) {
analysis::DefUseManager* def_use_mgr = context_->get_def_use_mgr();
CFG& cfg = *context_->cfg();
assert(cfg.preds(bb->id()).size() == 1 && "More than one predecessor");
std::unique_ptr<BasicBlock> new_bb =
MakeUnique<BasicBlock>(std::unique_ptr<Instruction>(new Instruction(
context_, SpvOpLabel, 0, context_->TakeNextId(), {})));
new_bb->SetParent(loop_utils_.GetFunction());
// Update the loop descriptor.
Loop* in_loop = (*loop_utils_.GetLoopDescriptor())[bb];
if (in_loop) {
in_loop->AddBasicBlock(new_bb.get());
loop_utils_.GetLoopDescriptor()->SetBasicBlockToLoop(new_bb->id(), in_loop);
}
context_->set_instr_block(new_bb->GetLabelInst(), new_bb.get());
def_use_mgr->AnalyzeInstDefUse(new_bb->GetLabelInst());
BasicBlock* bb_pred = cfg.block(cfg.preds(bb->id())[0]);
bb_pred->tail()->ForEachInId([bb, &new_bb](uint32_t* id) {
if (*id == bb->id()) {
*id = new_bb->id();
}
});
cfg.RemoveEdge(bb_pred->id(), bb->id());
cfg.AddEdge(bb_pred->id(), new_bb->id());
def_use_mgr->AnalyzeInstUse(&*bb_pred->tail());
// Update the incoming branch.
bb->ForEachPhiInst([&new_bb, def_use_mgr](Instruction* phi) {
phi->SetInOperand(1, {new_bb->id()});
def_use_mgr->AnalyzeInstUse(phi);
});
InstructionBuilder(
context_, new_bb.get(),
IRContext::kAnalysisDefUse | IRContext::kAnalysisInstrToBlockMapping)
.AddBranch(bb->id());
cfg.RegisterBlock(new_bb.get());
// Add the basic block to the function.
Function::iterator it = loop_utils_.GetFunction()->FindBlock(bb->id());
assert(it != loop_utils_.GetFunction()->end() &&
"Basic block not found in the function.");
BasicBlock* ret = new_bb.get();
loop_utils_.GetFunction()->AddBasicBlock(std::move(new_bb), it);
return ret;
}
BasicBlock* LoopPeeling::ProtectLoop(Loop* loop, Instruction* condition,
BasicBlock* if_merge) {
BasicBlock* if_block = loop->GetOrCreatePreHeaderBlock();
// Will no longer be a pre-header because of the if.
loop->SetPreHeaderBlock(nullptr);
// Kill the branch to the header.
context_->KillInst(&*if_block->tail());
InstructionBuilder builder(
context_, if_block,
IRContext::kAnalysisDefUse | IRContext::kAnalysisInstrToBlockMapping);
builder.AddConditionalBranch(condition->result_id(),
loop->GetHeaderBlock()->id(), if_merge->id(),
if_merge->id());
return if_block;
}
void LoopPeeling::PeelBefore(uint32_t peel_factor) {
assert(CanPeelLoop() && "Cannot peel loop");
LoopUtils::LoopCloningResult clone_results;
// Clone the loop and insert the cloned one before the loop.
DuplicateAndConnectLoop(&clone_results);
// Add a canonical induction variable "canonical_induction_variable_".
InsertCanonicalInductionVariable(&clone_results);
InstructionBuilder builder(
context_, &*cloned_loop_->GetPreHeaderBlock()->tail(),
IRContext::kAnalysisDefUse | IRContext::kAnalysisInstrToBlockMapping);
Instruction* factor =
builder.Add32BitConstantInteger(peel_factor, int_type_->IsSigned());
Instruction* has_remaining_iteration = builder.AddLessThan(
factor->result_id(), loop_iteration_count_->result_id());
Instruction* max_iteration = builder.AddSelect(
factor->type_id(), has_remaining_iteration->result_id(),
factor->result_id(), loop_iteration_count_->result_id());
// Change the exit condition of the cloned loop to be (exit when become
// false):
// "canonical_induction_variable_" < min("factor", "loop_iteration_count_")
FixExitCondition([max_iteration, this](Instruction* insert_before_point) {
return InstructionBuilder(context_, insert_before_point,
IRContext::kAnalysisDefUse |
IRContext::kAnalysisInstrToBlockMapping)
.AddLessThan(canonical_induction_variable_->result_id(),
max_iteration->result_id())
->result_id();
});
// "Protect" the second loop: the second loop can only be executed if
// |has_remaining_iteration| is true (i.e. factor < loop_iteration_count_).
BasicBlock* if_merge_block = loop_->GetMergeBlock();
loop_->SetMergeBlock(CreateBlockBefore(loop_->GetMergeBlock()));
// Prevent the second loop from being executed if we already executed all the
// required iterations.
BasicBlock* if_block =
ProtectLoop(loop_, has_remaining_iteration, if_merge_block);
// Patch the phi of the merge block.
if_merge_block->ForEachPhiInst(
[&clone_results, if_block, this](Instruction* phi) {
// if_merge_block had previously only 1 predecessor.
uint32_t incoming_value = phi->GetSingleWordInOperand(0);
auto def_in_loop = clone_results.value_map_.find(incoming_value);
if (def_in_loop != clone_results.value_map_.end())
incoming_value = def_in_loop->second;
phi->AddOperand(
{spv_operand_type_t::SPV_OPERAND_TYPE_ID, {incoming_value}});
phi->AddOperand(
{spv_operand_type_t::SPV_OPERAND_TYPE_ID, {if_block->id()}});
context_->get_def_use_mgr()->AnalyzeInstUse(phi);
});
context_->InvalidateAnalysesExceptFor(
IRContext::kAnalysisDefUse | IRContext::kAnalysisInstrToBlockMapping |
IRContext::kAnalysisLoopAnalysis | IRContext::kAnalysisCFG);
}
void LoopPeeling::PeelAfter(uint32_t peel_factor) {
assert(CanPeelLoop() && "Cannot peel loop");
LoopUtils::LoopCloningResult clone_results;
// Clone the loop and insert the cloned one before the loop.
DuplicateAndConnectLoop(&clone_results);
// Add a canonical induction variable "canonical_induction_variable_".
InsertCanonicalInductionVariable(&clone_results);
InstructionBuilder builder(
context_, &*cloned_loop_->GetPreHeaderBlock()->tail(),
IRContext::kAnalysisDefUse | IRContext::kAnalysisInstrToBlockMapping);
Instruction* factor =
builder.Add32BitConstantInteger(peel_factor, int_type_->IsSigned());
Instruction* has_remaining_iteration = builder.AddLessThan(
factor->result_id(), loop_iteration_count_->result_id());
// Change the exit condition of the cloned loop to be (exit when become
// false):
// "canonical_induction_variable_" + "factor" < "loop_iteration_count_"
FixExitCondition([factor, this](Instruction* insert_before_point) {
InstructionBuilder cond_builder(
context_, insert_before_point,
IRContext::kAnalysisDefUse | IRContext::kAnalysisInstrToBlockMapping);
// Build the following check: canonical_induction_variable_ + factor <
// iteration_count
return cond_builder
.AddLessThan(cond_builder
.AddIAdd(canonical_induction_variable_->type_id(),
canonical_induction_variable_->result_id(),
factor->result_id())
->result_id(),
loop_iteration_count_->result_id())
->result_id();
});
// "Protect" the first loop: the first loop can only be executed if
// factor < loop_iteration_count_.
// The original loop's pre-header was the cloned loop merge block.
GetClonedLoop()->SetMergeBlock(
CreateBlockBefore(GetOriginalLoop()->GetPreHeaderBlock()));
// Use the second loop preheader as if merge block.
// Prevent the first loop if only the peeled loop needs it.
BasicBlock* if_block = ProtectLoop(cloned_loop_, has_remaining_iteration,
GetOriginalLoop()->GetPreHeaderBlock());
// Patch the phi of the header block.
// We added an if to enclose the first loop and because the phi node are
// connected to the exit value of the first loop, the definition no longer
// dominate the preheader.
// We had to the preheader (our if merge block) the required phi instruction
// and patch the header phi.
GetOriginalLoop()->GetHeaderBlock()->ForEachPhiInst(
[&clone_results, if_block, this](Instruction* phi) {
analysis::DefUseManager* def_use_mgr = context_->get_def_use_mgr();
auto find_value_idx = [](Instruction* phi_inst, Loop* loop) {
uint32_t preheader_value_idx =
!loop->IsInsideLoop(phi_inst->GetSingleWordInOperand(1)) ? 0 : 2;
return preheader_value_idx;
};
Instruction* cloned_phi =
def_use_mgr->GetDef(clone_results.value_map_.at(phi->result_id()));
uint32_t cloned_preheader_value = cloned_phi->GetSingleWordInOperand(
find_value_idx(cloned_phi, GetClonedLoop()));
Instruction* new_phi =
InstructionBuilder(context_,
&*GetOriginalLoop()->GetPreHeaderBlock()->tail(),
IRContext::kAnalysisDefUse |
IRContext::kAnalysisInstrToBlockMapping)
.AddPhi(phi->type_id(),
{phi->GetSingleWordInOperand(
find_value_idx(phi, GetOriginalLoop())),
GetClonedLoop()->GetMergeBlock()->id(),
cloned_preheader_value, if_block->id()});
phi->SetInOperand(find_value_idx(phi, GetOriginalLoop()),
{new_phi->result_id()});
def_use_mgr->AnalyzeInstUse(phi);
});
context_->InvalidateAnalysesExceptFor(
IRContext::kAnalysisDefUse | IRContext::kAnalysisInstrToBlockMapping |
IRContext::kAnalysisLoopAnalysis | IRContext::kAnalysisCFG);
}
Pass::Status LoopPeelingPass::Process() {
bool modified = false;
Module* module = context()->module();
// Process each function in the module
for (Function& f : *module) {
modified |= ProcessFunction(&f);
}
return modified ? Status::SuccessWithChange : Status::SuccessWithoutChange;
}
bool LoopPeelingPass::ProcessFunction(Function* f) {
bool modified = false;
LoopDescriptor& loop_descriptor = *context()->GetLoopDescriptor(f);
std::vector<Loop*> to_process_loop;
to_process_loop.reserve(loop_descriptor.NumLoops());
for (Loop& l : loop_descriptor) {
to_process_loop.push_back(&l);
}
ScalarEvolutionAnalysis scev_analysis(context());
for (Loop* loop : to_process_loop) {
CodeMetrics loop_size;
loop_size.Analyze(*loop);
auto try_peel = [&loop_size, &modified, this](Loop* loop_to_peel) -> Loop* {
if (!loop_to_peel->IsLCSSA()) {
LoopUtils(context(), loop_to_peel).MakeLoopClosedSSA();
}
bool peeled_loop;
Loop* still_peelable_loop;
std::tie(peeled_loop, still_peelable_loop) =
ProcessLoop(loop_to_peel, &loop_size);
if (peeled_loop) {
modified = true;
}
return still_peelable_loop;
};
Loop* still_peelable_loop = try_peel(loop);
// The pass is working out the maximum factor by which a loop can be peeled.
// If the loop can potentially be peeled again, then there is only one
// possible direction, so only one call is still needed.
if (still_peelable_loop) {
try_peel(loop);
}
}
return modified;
}
std::pair<bool, Loop*> LoopPeelingPass::ProcessLoop(Loop* loop,
CodeMetrics* loop_size) {
ScalarEvolutionAnalysis* scev_analysis =
context()->GetScalarEvolutionAnalysis();
// Default values for bailing out.
std::pair<bool, Loop*> bail_out{false, nullptr};
BasicBlock* exit_block = loop->FindConditionBlock();
if (!exit_block) {
return bail_out;
}
Instruction* exiting_iv = loop->FindConditionVariable(exit_block);
if (!exiting_iv) {
return bail_out;
}
size_t iterations = 0;
if (!loop->FindNumberOfIterations(exiting_iv, &*exit_block->tail(),
&iterations)) {
return bail_out;
}
if (!iterations) {
return bail_out;
}
Instruction* canonical_induction_variable = nullptr;
loop->GetHeaderBlock()->WhileEachPhiInst([&canonical_induction_variable,
scev_analysis,
this](Instruction* insn) {
if (const SERecurrentNode* iv =
scev_analysis->AnalyzeInstruction(insn)->AsSERecurrentNode()) {
const SEConstantNode* offset = iv->GetOffset()->AsSEConstantNode();
const SEConstantNode* coeff = iv->GetCoefficient()->AsSEConstantNode();
if (offset && coeff && offset->FoldToSingleValue() == 0 &&
coeff->FoldToSingleValue() == 1) {
if (context()->get_type_mgr()->GetType(insn->type_id())->AsInteger()) {
canonical_induction_variable = insn;
return false;
}
}
}
return true;
});
bool is_signed = canonical_induction_variable
? context()
->get_type_mgr()
->GetType(canonical_induction_variable->type_id())
->AsInteger()
->IsSigned()
: false;
LoopPeeling peeler(
loop,
InstructionBuilder(
context(), loop->GetHeaderBlock(),
IRContext::kAnalysisDefUse | IRContext::kAnalysisInstrToBlockMapping)
.Add32BitConstantInteger<uint32_t>(static_cast<uint32_t>(iterations),
is_signed),
canonical_induction_variable);
if (!peeler.CanPeelLoop()) {
return bail_out;
}
// For each basic block in the loop, check if it can be peeled. If it
// can, get the direction (before/after) and by which factor.
LoopPeelingInfo peel_info(loop, iterations, scev_analysis);
uint32_t peel_before_factor = 0;
uint32_t peel_after_factor = 0;
for (uint32_t block : loop->GetBlocks()) {
if (block == exit_block->id()) {
continue;
}
BasicBlock* bb = cfg()->block(block);
PeelDirection direction;
uint32_t factor;
std::tie(direction, factor) = peel_info.GetPeelingInfo(bb);
if (direction == PeelDirection::kNone) {
continue;
}
if (direction == PeelDirection::kBefore) {
peel_before_factor = std::max(peel_before_factor, factor);
} else {
assert(direction == PeelDirection::kAfter);
peel_after_factor = std::max(peel_after_factor, factor);
}
}
PeelDirection direction = PeelDirection::kNone;
uint32_t factor = 0;
// Find which direction we should peel.
if (peel_before_factor) {
factor = peel_before_factor;
direction = PeelDirection::kBefore;
}
if (peel_after_factor) {
if (peel_before_factor < peel_after_factor) {
// Favor a peel after here and give the peel before another shot later.
factor = peel_after_factor;
direction = PeelDirection::kAfter;
}
}
// Do the peel if we can.
if (direction == PeelDirection::kNone) return bail_out;
// This does not take into account branch elimination opportunities and
// the unrolling. It assumes the peeled loop will be unrolled as well.
if (factor * loop_size->roi_size_ > code_grow_threshold_) {
return bail_out;
}
loop_size->roi_size_ *= factor;
// Find if a loop should be peeled again.
Loop* extra_opportunity = nullptr;
if (direction == PeelDirection::kBefore) {
peeler.PeelBefore(factor);
if (stats_) {
stats_->peeled_loops_.emplace_back(loop, PeelDirection::kBefore, factor);
}
if (peel_after_factor) {
// We could have peeled after, give it another try.
extra_opportunity = peeler.GetOriginalLoop();
}
} else {
peeler.PeelAfter(factor);
if (stats_) {
stats_->peeled_loops_.emplace_back(loop, PeelDirection::kAfter, factor);
}
if (peel_before_factor) {
// We could have peeled before, give it another try.
extra_opportunity = peeler.GetClonedLoop();
}
}
return {true, extra_opportunity};
}
uint32_t LoopPeelingPass::LoopPeelingInfo::GetFirstLoopInvariantOperand(
Instruction* condition) const {
for (uint32_t i = 0; i < condition->NumInOperands(); i++) {
BasicBlock* bb =
context_->get_instr_block(condition->GetSingleWordInOperand(i));
if (bb && loop_->IsInsideLoop(bb)) {
return condition->GetSingleWordInOperand(i);
}
}
return 0;
}
uint32_t LoopPeelingPass::LoopPeelingInfo::GetFirstNonLoopInvariantOperand(
Instruction* condition) const {
for (uint32_t i = 0; i < condition->NumInOperands(); i++) {
BasicBlock* bb =
context_->get_instr_block(condition->GetSingleWordInOperand(i));
if (!bb || !loop_->IsInsideLoop(bb)) {
return condition->GetSingleWordInOperand(i);
}
}
return 0;
}
static bool IsHandledCondition(SpvOp opcode) {
switch (opcode) {
case SpvOpIEqual:
case SpvOpINotEqual:
case SpvOpUGreaterThan:
case SpvOpSGreaterThan:
case SpvOpUGreaterThanEqual:
case SpvOpSGreaterThanEqual:
case SpvOpULessThan:
case SpvOpSLessThan:
case SpvOpULessThanEqual:
case SpvOpSLessThanEqual:
return true;
default:
return false;
}
}
LoopPeelingPass::LoopPeelingInfo::Direction
LoopPeelingPass::LoopPeelingInfo::GetPeelingInfo(BasicBlock* bb) const {
if (bb->terminator()->opcode() != SpvOpBranchConditional) {
return GetNoneDirection();
}
analysis::DefUseManager* def_use_mgr = context_->get_def_use_mgr();
Instruction* condition =
def_use_mgr->GetDef(bb->terminator()->GetSingleWordInOperand(0));
if (!IsHandledCondition(condition->opcode())) {
return GetNoneDirection();
}
if (!GetFirstLoopInvariantOperand(condition)) {
// No loop invariant, it cannot be peeled by this pass.
return GetNoneDirection();
}
if (!GetFirstNonLoopInvariantOperand(condition)) {
// Seems to be a job for the unswitch pass.
return GetNoneDirection();
}
// Left hand-side.
SExpression lhs = scev_analysis_->AnalyzeInstruction(
def_use_mgr->GetDef(condition->GetSingleWordInOperand(0)));
if (lhs->GetType() == SENode::CanNotCompute) {
// Can't make any conclusion.
return GetNoneDirection();
}
// Right hand-side.
SExpression rhs = scev_analysis_->AnalyzeInstruction(
def_use_mgr->GetDef(condition->GetSingleWordInOperand(1)));
if (rhs->GetType() == SENode::CanNotCompute) {
// Can't make any conclusion.
return GetNoneDirection();
}
// Only take into account recurrent expression over the current loop.
bool is_lhs_rec = !scev_analysis_->IsLoopInvariant(loop_, lhs);
bool is_rhs_rec = !scev_analysis_->IsLoopInvariant(loop_, rhs);
if ((is_lhs_rec && is_rhs_rec) || (!is_lhs_rec && !is_rhs_rec)) {
return GetNoneDirection();
}
if (is_lhs_rec) {
if (!lhs->AsSERecurrentNode() ||
lhs->AsSERecurrentNode()->GetLoop() != loop_) {
return GetNoneDirection();
}
}
if (is_rhs_rec) {
if (!rhs->AsSERecurrentNode() ||
rhs->AsSERecurrentNode()->GetLoop() != loop_) {
return GetNoneDirection();
}
}
// If the op code is ==, then we try a peel before or after.
// If opcode is not <, >, <= or >=, we bail out.
//
// For the remaining cases, we canonicalize the expression so that the
// constant expression is on the left hand side and the recurring expression
// is on the right hand side. If we swap hand side, then < becomes >, <=
// becomes >= etc.
// If the opcode is <=, then we add 1 to the right hand side and do the peel
// check on <.
// If the opcode is >=, then we add 1 to the left hand side and do the peel
// check on >.
CmpOperator cmp_operator;
switch (condition->opcode()) {
default:
return GetNoneDirection();
case SpvOpIEqual:
case SpvOpINotEqual:
return HandleEquality(lhs, rhs);
case SpvOpUGreaterThan:
case SpvOpSGreaterThan: {
cmp_operator = CmpOperator::kGT;
break;
}
case SpvOpULessThan:
case SpvOpSLessThan: {
cmp_operator = CmpOperator::kLT;
break;
}
// We add one to transform >= into > and <= into <.
case SpvOpUGreaterThanEqual:
case SpvOpSGreaterThanEqual: {
cmp_operator = CmpOperator::kGE;
break;
}
case SpvOpULessThanEqual:
case SpvOpSLessThanEqual: {
cmp_operator = CmpOperator::kLE;
break;
}
}
// Force the left hand side to be the non recurring expression.
if (is_lhs_rec) {
std::swap(lhs, rhs);
switch (cmp_operator) {
case CmpOperator::kLT: {
cmp_operator = CmpOperator::kGT;
break;
}
case CmpOperator::kGT: {
cmp_operator = CmpOperator::kLT;
break;
}
case CmpOperator::kLE: {
cmp_operator = CmpOperator::kGE;
break;
}
case CmpOperator::kGE: {
cmp_operator = CmpOperator::kLE;
break;
}
}
}
return HandleInequality(cmp_operator, lhs, rhs->AsSERecurrentNode());
}
SExpression LoopPeelingPass::LoopPeelingInfo::GetValueAtFirstIteration(
SERecurrentNode* rec) const {
return rec->GetOffset();
}
SExpression LoopPeelingPass::LoopPeelingInfo::GetValueAtIteration(
SERecurrentNode* rec, int64_t iteration) const {
SExpression coeff = rec->GetCoefficient();
SExpression offset = rec->GetOffset();
return (coeff * iteration) + offset;
}
SExpression LoopPeelingPass::LoopPeelingInfo::GetValueAtLastIteration(
SERecurrentNode* rec) const {
return GetValueAtIteration(rec, loop_max_iterations_ - 1);
}
bool LoopPeelingPass::LoopPeelingInfo::EvalOperator(CmpOperator cmp_op,
SExpression lhs,
SExpression rhs,
bool* result) const {
assert(scev_analysis_->IsLoopInvariant(loop_, lhs));
assert(scev_analysis_->IsLoopInvariant(loop_, rhs));
// We perform the test: 0 cmp_op rhs - lhs
// What is left is then to determine the sign of the expression.
switch (cmp_op) {
case CmpOperator::kLT: {
return scev_analysis_->IsAlwaysGreaterThanZero(rhs - lhs, result);
}
case CmpOperator::kGT: {
return scev_analysis_->IsAlwaysGreaterThanZero(lhs - rhs, result);
}
case CmpOperator::kLE: {
return scev_analysis_->IsAlwaysGreaterOrEqualToZero(rhs - lhs, result);
}
case CmpOperator::kGE: {
return scev_analysis_->IsAlwaysGreaterOrEqualToZero(lhs - rhs, result);
}
}
return false;
}
LoopPeelingPass::LoopPeelingInfo::Direction
LoopPeelingPass::LoopPeelingInfo::HandleEquality(SExpression lhs,
SExpression rhs) const {
{
// Try peel before opportunity.
SExpression lhs_cst = lhs;
if (SERecurrentNode* rec_node = lhs->AsSERecurrentNode()) {
lhs_cst = rec_node->GetOffset();
}
SExpression rhs_cst = rhs;
if (SERecurrentNode* rec_node = rhs->AsSERecurrentNode()) {
rhs_cst = rec_node->GetOffset();
}
if (lhs_cst == rhs_cst) {
return Direction{LoopPeelingPass::PeelDirection::kBefore, 1};
}
}
{
// Try peel after opportunity.
SExpression lhs_cst = lhs;
if (SERecurrentNode* rec_node = lhs->AsSERecurrentNode()) {
// rec_node(x) = a * x + b
// assign to lhs: a * (loop_max_iterations_ - 1) + b
lhs_cst = GetValueAtLastIteration(rec_node);
}
SExpression rhs_cst = rhs;
if (SERecurrentNode* rec_node = rhs->AsSERecurrentNode()) {
// rec_node(x) = a * x + b
// assign to lhs: a * (loop_max_iterations_ - 1) + b
rhs_cst = GetValueAtLastIteration(rec_node);
}
if (lhs_cst == rhs_cst) {
return Direction{LoopPeelingPass::PeelDirection::kAfter, 1};
}
}
return GetNoneDirection();
}
LoopPeelingPass::LoopPeelingInfo::Direction
LoopPeelingPass::LoopPeelingInfo::HandleInequality(CmpOperator cmp_op,
SExpression lhs,
SERecurrentNode* rhs) const {
SExpression offset = rhs->GetOffset();
SExpression coefficient = rhs->GetCoefficient();
// Compute (cst - B) / A.
std::pair<SExpression, int64_t> flip_iteration = (lhs - offset) / coefficient;
if (!flip_iteration.first->AsSEConstantNode()) {
return GetNoneDirection();
}
// note: !!flip_iteration.second normalize to 0/1 (via bool cast).
int64_t iteration =
flip_iteration.first->AsSEConstantNode()->FoldToSingleValue() +
!!flip_iteration.second;
if (iteration <= 0 ||
loop_max_iterations_ <= static_cast<uint64_t>(iteration)) {
// Always true or false within the loop bounds.
return GetNoneDirection();
}
// If this is a <= or >= operator and the iteration, make sure |iteration| is
// the one flipping the condition.
// If (cst - B) and A are not divisible, this equivalent to a < or > check, so
// we skip this test.
if (!flip_iteration.second &&
(cmp_op == CmpOperator::kLE || cmp_op == CmpOperator::kGE)) {
bool first_iteration;
bool current_iteration;
if (!EvalOperator(cmp_op, lhs, offset, &first_iteration) ||
!EvalOperator(cmp_op, lhs, GetValueAtIteration(rhs, iteration),
&current_iteration)) {
return GetNoneDirection();
}
// If the condition did not flip the next will.
if (first_iteration == current_iteration) {
iteration++;
}
}
uint32_t cast_iteration = 0;
// sanity check: can we fit |iteration| in a uint32_t ?
if (static_cast<uint64_t>(iteration) < std::numeric_limits<uint32_t>::max()) {
cast_iteration = static_cast<uint32_t>(iteration);
}
if (cast_iteration) {
// Peel before if we are closer to the start, after if closer to the end.
if (loop_max_iterations_ / 2 > cast_iteration) {
return Direction{LoopPeelingPass::PeelDirection::kBefore, cast_iteration};
} else {
return Direction{
LoopPeelingPass::PeelDirection::kAfter,
static_cast<uint32_t>(loop_max_iterations_ - cast_iteration)};
}
}
return GetNoneDirection();
}
} // namespace opt
} // namespace spvtools