mirror of
https://github.com/KhronosGroup/SPIRV-Tools
synced 2024-12-02 23:50:13 +00:00
7075c49923
The current implementation of merge return can create bad, but correct, code. When it is not in a loop construct, it will insert a lot of extra branch around code. The potentially large number of branches are bad. At the same time, it can separate code store to variables from its uses hiding the fact that the store dominates the load. This hurts the later analysis because the compiler thinks that multiple values can reach a load, when there is really only 1. This poorer analysis leads to missed optimizations. The solution is to create a dummy loop around the entire body of the function, then we can break from that loop with a single branch. Also only new merge nodes would be those at the end of loops meaning that most analysies will not be hurt. Remove dead code for cases that are no longer possible. It seems like some drivers expect there the be an OpSelectionMerge before conditional branches, even if they are not strictly needed. So we add them.
314 lines
11 KiB
C++
314 lines
11 KiB
C++
// Copyright (c) 2017 Google Inc.
|
|
//
|
|
// 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/opt/cfg.h"
|
|
|
|
#include <memory>
|
|
#include <utility>
|
|
|
|
#include "source/cfa.h"
|
|
#include "source/opt/ir_builder.h"
|
|
#include "source/opt/ir_context.h"
|
|
#include "source/opt/module.h"
|
|
|
|
namespace spvtools {
|
|
namespace opt {
|
|
namespace {
|
|
|
|
using cbb_ptr = const opt::BasicBlock*;
|
|
|
|
// Universal Limit of ResultID + 1
|
|
const int kMaxResultId = 0x400000;
|
|
|
|
} // namespace
|
|
|
|
CFG::CFG(Module* module)
|
|
: module_(module),
|
|
pseudo_entry_block_(std::unique_ptr<Instruction>(
|
|
new Instruction(module->context(), SpvOpLabel, 0, 0, {}))),
|
|
pseudo_exit_block_(std::unique_ptr<Instruction>(new Instruction(
|
|
module->context(), SpvOpLabel, 0, kMaxResultId, {}))) {
|
|
for (auto& fn : *module) {
|
|
for (auto& blk : fn) {
|
|
RegisterBlock(&blk);
|
|
}
|
|
}
|
|
}
|
|
|
|
void CFG::AddEdges(BasicBlock* blk) {
|
|
uint32_t blk_id = blk->id();
|
|
// Force the creation of an entry, not all basic block have predecessors
|
|
// (such as the entry blocks and some unreachables).
|
|
label2preds_[blk_id];
|
|
const auto* const_blk = blk;
|
|
const_blk->ForEachSuccessorLabel(
|
|
[blk_id, this](const uint32_t succ_id) { AddEdge(blk_id, succ_id); });
|
|
}
|
|
|
|
void CFG::RemoveNonExistingEdges(uint32_t blk_id) {
|
|
std::vector<uint32_t> updated_pred_list;
|
|
for (uint32_t id : preds(blk_id)) {
|
|
const BasicBlock* pred_blk = block(id);
|
|
bool has_branch = false;
|
|
pred_blk->ForEachSuccessorLabel([&has_branch, blk_id](uint32_t succ) {
|
|
if (succ == blk_id) {
|
|
has_branch = true;
|
|
}
|
|
});
|
|
if (has_branch) updated_pred_list.push_back(id);
|
|
}
|
|
|
|
label2preds_.at(blk_id) = std::move(updated_pred_list);
|
|
}
|
|
|
|
void CFG::ComputeStructuredOrder(Function* func, BasicBlock* root,
|
|
std::list<BasicBlock*>* order) {
|
|
assert(module_->context()->get_feature_mgr()->HasCapability(
|
|
SpvCapabilityShader) &&
|
|
"This only works on structured control flow");
|
|
|
|
// Compute structured successors and do DFS.
|
|
ComputeStructuredSuccessors(func);
|
|
auto ignore_block = [](cbb_ptr) {};
|
|
auto ignore_edge = [](cbb_ptr, cbb_ptr) {};
|
|
auto get_structured_successors = [this](const BasicBlock* b) {
|
|
return &(block2structured_succs_[b]);
|
|
};
|
|
|
|
// TODO(greg-lunarg): Get rid of const_cast by making moving const
|
|
// out of the cfa.h prototypes and into the invoking code.
|
|
auto post_order = [&](cbb_ptr b) {
|
|
order->push_front(const_cast<BasicBlock*>(b));
|
|
};
|
|
CFA<BasicBlock>::DepthFirstTraversal(root, get_structured_successors,
|
|
ignore_block, post_order, ignore_edge);
|
|
}
|
|
|
|
void CFG::ForEachBlockInPostOrder(BasicBlock* bb,
|
|
const std::function<void(BasicBlock*)>& f) {
|
|
std::vector<BasicBlock*> po;
|
|
std::unordered_set<BasicBlock*> seen;
|
|
ComputePostOrderTraversal(bb, &po, &seen);
|
|
|
|
for (BasicBlock* current_bb : po) {
|
|
if (!IsPseudoExitBlock(current_bb) && !IsPseudoEntryBlock(current_bb)) {
|
|
f(current_bb);
|
|
}
|
|
}
|
|
}
|
|
|
|
void CFG::ForEachBlockInReversePostOrder(
|
|
BasicBlock* bb, const std::function<void(BasicBlock*)>& f) {
|
|
std::vector<BasicBlock*> po;
|
|
std::unordered_set<BasicBlock*> seen;
|
|
ComputePostOrderTraversal(bb, &po, &seen);
|
|
|
|
for (auto current_bb = po.rbegin(); current_bb != po.rend(); ++current_bb) {
|
|
if (!IsPseudoExitBlock(*current_bb) && !IsPseudoEntryBlock(*current_bb)) {
|
|
f(*current_bb);
|
|
}
|
|
}
|
|
}
|
|
|
|
void CFG::ComputeStructuredSuccessors(Function* func) {
|
|
block2structured_succs_.clear();
|
|
for (auto& blk : *func) {
|
|
// If no predecessors in function, make successor to pseudo entry.
|
|
if (label2preds_[blk.id()].size() == 0)
|
|
block2structured_succs_[&pseudo_entry_block_].push_back(&blk);
|
|
|
|
// If header, make merge block first successor and continue block second
|
|
// successor if there is one.
|
|
uint32_t mbid = blk.MergeBlockIdIfAny();
|
|
if (mbid != 0) {
|
|
block2structured_succs_[&blk].push_back(block(mbid));
|
|
uint32_t cbid = blk.ContinueBlockIdIfAny();
|
|
if (cbid != 0) {
|
|
block2structured_succs_[&blk].push_back(block(cbid));
|
|
}
|
|
}
|
|
|
|
// Add true successors.
|
|
const auto& const_blk = blk;
|
|
const_blk.ForEachSuccessorLabel([&blk, this](const uint32_t sbid) {
|
|
block2structured_succs_[&blk].push_back(block(sbid));
|
|
});
|
|
}
|
|
}
|
|
|
|
void CFG::ComputePostOrderTraversal(BasicBlock* bb,
|
|
std::vector<BasicBlock*>* order,
|
|
std::unordered_set<BasicBlock*>* seen) {
|
|
seen->insert(bb);
|
|
static_cast<const BasicBlock*>(bb)->ForEachSuccessorLabel(
|
|
[&order, &seen, this](const uint32_t sbid) {
|
|
BasicBlock* succ_bb = id2block_[sbid];
|
|
if (!seen->count(succ_bb)) {
|
|
ComputePostOrderTraversal(succ_bb, order, seen);
|
|
}
|
|
});
|
|
order->push_back(bb);
|
|
}
|
|
|
|
BasicBlock* CFG::SplitLoopHeader(BasicBlock* bb) {
|
|
assert(bb->GetLoopMergeInst() && "Expecting bb to be the header of a loop.");
|
|
|
|
Function* fn = bb->GetParent();
|
|
IRContext* context = module_->context();
|
|
|
|
// Find the insertion point for the new bb.
|
|
Function::iterator header_it = std::find_if(
|
|
fn->begin(), fn->end(),
|
|
[bb](BasicBlock& block_in_func) { return &block_in_func == bb; });
|
|
assert(header_it != fn->end());
|
|
|
|
const std::vector<uint32_t>& pred = preds(bb->id());
|
|
// Find the back edge
|
|
BasicBlock* latch_block = nullptr;
|
|
Function::iterator latch_block_iter = header_it;
|
|
while (++latch_block_iter != fn->end()) {
|
|
// If blocks are in the proper order, then the only branch that appears
|
|
// after the header is the latch.
|
|
if (std::find(pred.begin(), pred.end(), latch_block_iter->id()) !=
|
|
pred.end()) {
|
|
break;
|
|
}
|
|
}
|
|
assert(latch_block_iter != fn->end() && "Could not find the latch.");
|
|
latch_block = &*latch_block_iter;
|
|
|
|
RemoveSuccessorEdges(bb);
|
|
|
|
// Create the new header bb basic bb.
|
|
// Leave the phi instructions behind.
|
|
auto iter = bb->begin();
|
|
while (iter->opcode() == SpvOpPhi) {
|
|
++iter;
|
|
}
|
|
|
|
BasicBlock* new_header =
|
|
bb->SplitBasicBlock(context, context->TakeNextId(), iter);
|
|
|
|
uint32_t new_header_id = new_header->id();
|
|
context->AnalyzeDefUse(new_header->GetLabelInst());
|
|
|
|
// Update cfg
|
|
RegisterBlock(new_header);
|
|
|
|
// Update bb mappings.
|
|
context->set_instr_block(new_header->GetLabelInst(), new_header);
|
|
new_header->ForEachInst([new_header, context](Instruction* inst) {
|
|
context->set_instr_block(inst, new_header);
|
|
});
|
|
|
|
// Adjust the OpPhi instructions as needed.
|
|
bb->ForEachPhiInst([latch_block, bb, new_header, context](Instruction* phi) {
|
|
std::vector<uint32_t> preheader_phi_ops;
|
|
std::vector<Operand> header_phi_ops;
|
|
|
|
// Identify where the original inputs to original OpPhi belong: header or
|
|
// preheader.
|
|
for (uint32_t i = 0; i < phi->NumInOperands(); i += 2) {
|
|
uint32_t def_id = phi->GetSingleWordInOperand(i);
|
|
uint32_t branch_id = phi->GetSingleWordInOperand(i + 1);
|
|
if (branch_id == latch_block->id()) {
|
|
header_phi_ops.push_back({SPV_OPERAND_TYPE_ID, {def_id}});
|
|
header_phi_ops.push_back({SPV_OPERAND_TYPE_ID, {branch_id}});
|
|
} else {
|
|
preheader_phi_ops.push_back(def_id);
|
|
preheader_phi_ops.push_back(branch_id);
|
|
}
|
|
}
|
|
|
|
// Create a phi instruction if and only if the preheader_phi_ops has more
|
|
// than one pair.
|
|
if (preheader_phi_ops.size() > 2) {
|
|
InstructionBuilder builder(
|
|
context, &*bb->begin(),
|
|
IRContext::kAnalysisDefUse | IRContext::kAnalysisInstrToBlockMapping);
|
|
|
|
Instruction* new_phi = builder.AddPhi(phi->type_id(), preheader_phi_ops);
|
|
|
|
// Add the OpPhi to the header bb.
|
|
header_phi_ops.push_back({SPV_OPERAND_TYPE_ID, {new_phi->result_id()}});
|
|
header_phi_ops.push_back({SPV_OPERAND_TYPE_ID, {bb->id()}});
|
|
} else {
|
|
// An OpPhi with a single entry is just a copy. In this case use the same
|
|
// instruction in the new header.
|
|
header_phi_ops.push_back({SPV_OPERAND_TYPE_ID, {preheader_phi_ops[0]}});
|
|
header_phi_ops.push_back({SPV_OPERAND_TYPE_ID, {bb->id()}});
|
|
}
|
|
|
|
phi->RemoveFromList();
|
|
std::unique_ptr<Instruction> phi_owner(phi);
|
|
phi->SetInOperands(std::move(header_phi_ops));
|
|
new_header->begin()->InsertBefore(std::move(phi_owner));
|
|
context->set_instr_block(phi, new_header);
|
|
context->AnalyzeUses(phi);
|
|
});
|
|
|
|
// Add a branch to the new header.
|
|
InstructionBuilder branch_builder(
|
|
context, bb,
|
|
IRContext::kAnalysisDefUse | IRContext::kAnalysisInstrToBlockMapping);
|
|
bb->AddInstruction(
|
|
MakeUnique<Instruction>(context, SpvOpBranch, 0, 0,
|
|
std::initializer_list<Operand>{
|
|
{SPV_OPERAND_TYPE_ID, {new_header->id()}}}));
|
|
context->AnalyzeUses(bb->terminator());
|
|
context->set_instr_block(bb->terminator(), bb);
|
|
label2preds_[new_header->id()].push_back(bb->id());
|
|
|
|
// Update the latch to branch to the new header.
|
|
latch_block->ForEachSuccessorLabel([bb, new_header_id](uint32_t* id) {
|
|
if (*id == bb->id()) {
|
|
*id = new_header_id;
|
|
}
|
|
});
|
|
Instruction* latch_branch = latch_block->terminator();
|
|
context->AnalyzeUses(latch_branch);
|
|
label2preds_[new_header->id()].push_back(latch_block->id());
|
|
|
|
auto& block_preds = label2preds_[bb->id()];
|
|
auto latch_pos =
|
|
std::find(block_preds.begin(), block_preds.end(), latch_block->id());
|
|
assert(latch_pos != block_preds.end() && "The cfg was invalid.");
|
|
block_preds.erase(latch_pos);
|
|
|
|
// Update the loop descriptors
|
|
if (context->AreAnalysesValid(IRContext::kAnalysisLoopAnalysis)) {
|
|
LoopDescriptor* loop_desc = context->GetLoopDescriptor(bb->GetParent());
|
|
Loop* loop = (*loop_desc)[bb->id()];
|
|
|
|
loop->AddBasicBlock(new_header_id);
|
|
loop->SetHeaderBlock(new_header);
|
|
loop_desc->SetBasicBlockToLoop(new_header_id, loop);
|
|
|
|
loop->RemoveBasicBlock(bb->id());
|
|
loop->SetPreHeaderBlock(bb);
|
|
|
|
Loop* parent_loop = loop->GetParent();
|
|
if (parent_loop != nullptr) {
|
|
parent_loop->AddBasicBlock(bb->id());
|
|
loop_desc->SetBasicBlockToLoop(bb->id(), parent_loop);
|
|
} else {
|
|
loop_desc->SetBasicBlockToLoop(bb->id(), nullptr);
|
|
}
|
|
}
|
|
return new_header;
|
|
}
|
|
|
|
} // namespace opt
|
|
} // namespace spvtools
|