AuroraMiddleware/SPIRV-Tools
1
0
Fork 0
mirror of https://github.com/KhronosGroup/SPIRV-Tools synced 2024-11-22 19:50:05 +00:00
Code Issues Packages Projects Releases Wiki Activity

mem2reg: Add LocalSingleStoreElimPass

Browse Source 
Eliminate function scope variables with one store, if possible.
This commit is contained in:
GregF 2017-05-19 17:31:28 -06:00 committed by David Neto
parent 7c8da66bc2
commit 0c5722fc01
9 changed files with 1243 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

18
include/spirv-tools/optimizer.hpp
View File

@ -231,6 +231,24 @@ Optimizer::PassToken CreateLocalSingleBlockLoadStoreElimPass();
// possible.
Optimizer::PassToken CreateLocalAccessChainConvertPass();
// Creates a local single store elimination pass.
// For each entry point function, this pass eliminates loads and stores for
// function scope variable that are stored to only once, where possible. Only
// whole variable loads and stores are eliminated; access-chain references are
// not optimized. Replace all loads of such variables with the value that is
// stored and eliminate any resulting dead code.
//
// Currently, the presence of access chains and function calls can inhibit this
// pass, however the Inlining and LocalAccessChainConvert passes can make it
// more effective. In additional, many non-load/store memory operations are
// not supported and will prohibit optimization of a function. Support of
// these operations are future work.
//
// This pass will reduce the work needed to be done by LocalSingleBlockElim
// and LocalSSARewrite and can improve the effectiveness of other passes such
// as DeadBranchElimination which depend on values for their analysis.
Optimizer::PassToken CreateLocalSingleStoreElimPass();
// Creates a compact ids pass.
// The pass remaps result ids to a compact and gapless range starting from %1.
Optimizer::PassToken CreateCompactIdsPass();

2
source/opt/CMakeLists.txt
View File

@ -27,6 +27,7 @@ add_library(SPIRV-Tools-opt
ir_loader.h
local_access_chain_convert_pass.h
local_single_block_elim_pass.h
local_single_store_elim_pass.h
log.h
module.h
null_pass.h
@ -54,6 +55,7 @@ add_library(SPIRV-Tools-opt
ir_loader.cpp
local_access_chain_convert_pass.cpp
local_single_block_elim_pass.cpp
local_single_store_elim_pass.cpp
module.cpp
set_spec_constant_default_value_pass.cpp
optimizer.cpp

469
source/opt/local_single_store_elim_pass.cpp Normal file
View File

@ -0,0 +1,469 @@
// Copyright (c) 2017 The Khronos Group Inc.
// Copyright (c) 2017 Valve Corporation
// Copyright (c) 2017 LunarG 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 "local_single_store_elim_pass.h"
#include "cfa.h"
#include "iterator.h"
#include "spirv/1.0/GLSL.std.450.h"
static const int kSpvEntryPointFunctionId = 1;
static const int kSpvStorePtrId = 0;
static const int kSpvStoreValId = 1;
static const int kSpvLoadPtrId = 0;
static const int kSpvAccessChainPtrId = 0;
static const int kSpvTypePointerStorageClass = 0;
static const int kSpvTypePointerTypeId = 1;
// Universal Limit of ResultID + 1
static const int kInvalidId = 0x400000;
namespace spvtools {
namespace opt {
bool LocalSingleStoreElimPass::IsNonPtrAccessChain(const SpvOp opcode) const {
return opcode == SpvOpAccessChain || opcode == SpvOpInBoundsAccessChain;
}
bool LocalSingleStoreElimPass::IsMathType(
const ir::Instruction* typeInst) const {
switch (typeInst->opcode()) {
case SpvOpTypeInt:
case SpvOpTypeFloat:
case SpvOpTypeBool:
case SpvOpTypeVector:
case SpvOpTypeMatrix:
return true;
default:
break;
}
return false;
}
bool LocalSingleStoreElimPass::IsTargetType(
const ir::Instruction* typeInst) const {
if (IsMathType(typeInst))
return true;
if (typeInst->opcode() == SpvOpTypeArray)
return IsMathType(def_use_mgr_->GetDef(typeInst->GetSingleWordOperand(1)));
if (typeInst->opcode() != SpvOpTypeStruct)
return false;
// All struct members must be math type
int nonMathComp = 0;
typeInst->ForEachInId([&nonMathComp,this](const uint32_t* tid) {
ir::Instruction* compTypeInst = def_use_mgr_->GetDef(*tid);
if (!IsMathType(compTypeInst)) ++nonMathComp;
});
return nonMathComp == 0;
}
ir::Instruction* LocalSingleStoreElimPass::GetPtr(
ir::Instruction* ip, uint32_t* varId) {
*varId = ip->GetSingleWordInOperand(
ip->opcode() == SpvOpStore ? kSpvStorePtrId : kSpvLoadPtrId);
ir::Instruction* ptrInst = def_use_mgr_->GetDef(*varId);
ir::Instruction* varInst = ptrInst;
while (IsNonPtrAccessChain(varInst->opcode())) {
*varId = varInst->GetSingleWordInOperand(kSpvAccessChainPtrId);
varInst = def_use_mgr_->GetDef(*varId);
}
return ptrInst;
}
bool LocalSingleStoreElimPass::IsTargetVar(uint32_t varId) {
if (seen_non_target_vars_.find(varId) != seen_non_target_vars_.end())
return false;
if (seen_target_vars_.find(varId) != seen_target_vars_.end())
return true;
const ir::Instruction* varInst = def_use_mgr_->GetDef(varId);
assert(varInst->opcode() == SpvOpVariable);
const uint32_t varTypeId = varInst->type_id();
const ir::Instruction* varTypeInst = def_use_mgr_->GetDef(varTypeId);
if (varTypeInst->GetSingleWordInOperand(kSpvTypePointerStorageClass) !=
SpvStorageClassFunction) {
seen_non_target_vars_.insert(varId);
return false;
}
const uint32_t varPteTypeId =
varTypeInst->GetSingleWordInOperand(kSpvTypePointerTypeId);
ir::Instruction* varPteTypeInst = def_use_mgr_->GetDef(varPteTypeId);
if (!IsTargetType(varPteTypeInst)) {
seen_non_target_vars_.insert(varId);
return false;
}
seen_target_vars_.insert(varId);
return true;
}
bool LocalSingleStoreElimPass::HasOnlySupportedRefs(uint32_t ptrId) {
if (supported_ref_ptrs_.find(ptrId) != supported_ref_ptrs_.end())
return true;
analysis::UseList* uses = def_use_mgr_->GetUses(ptrId);
assert(uses != nullptr);
for (auto u : *uses) {
SpvOp op = u.inst->opcode();
if (IsNonPtrAccessChain(op)) {
if (!HasOnlySupportedRefs(u.inst->result_id()))
return false;
}
else if (op != SpvOpStore && op != SpvOpLoad && op != SpvOpName)
return false;
}
supported_ref_ptrs_.insert(ptrId);
return true;
}
void LocalSingleStoreElimPass::SingleStoreAnalyze(ir::Function* func) {
ssa_var2store_.clear();
non_ssa_vars_.clear();
store2idx_.clear();
store2blk_.clear();
for (auto bi = func->begin(); bi != func->end(); ++bi) {
uint32_t instIdx = 0;
for (auto ii = bi->begin(); ii != bi->end(); ++ii, ++instIdx) {
switch (ii->opcode()) {
case SpvOpStore: {
// Verify store variable is target type
uint32_t varId;
ir::Instruction* ptrInst = GetPtr(&*ii, &varId);
if (non_ssa_vars_.find(varId) != non_ssa_vars_.end())
continue;
if (!HasOnlySupportedRefs(varId)) {
non_ssa_vars_.insert(varId);
continue;
}
if (IsNonPtrAccessChain(ptrInst->opcode())) {
non_ssa_vars_.insert(varId);
ssa_var2store_.erase(varId);
continue;
}
// Verify target type and function storage class
if (!IsTargetVar(varId)) {
non_ssa_vars_.insert(varId);
continue;
}
// Ignore variables with multiple stores
if (ssa_var2store_.find(varId) != ssa_var2store_.end()) {
non_ssa_vars_.insert(varId);
ssa_var2store_.erase(varId);
continue;
}
// Remember pointer to variable's store and it's
// ordinal position in block
ssa_var2store_[varId] = &*ii;
store2idx_[&*ii] = instIdx;
store2blk_[&*ii] = &*bi;
} break;
default:
break;
} // switch
}
}
}
void LocalSingleStoreElimPass::ReplaceAndDeleteLoad(
ir::Instruction* loadInst, uint32_t replId) {
(void) def_use_mgr_->ReplaceAllUsesWith(loadInst->result_id(), replId);
DCEInst(loadInst);
}
LocalSingleStoreElimPass::GetBlocksFunction
LocalSingleStoreElimPass::AugmentedCFGSuccessorsFunction() const {
return [this](const ir::BasicBlock* block) {
auto asmi = augmented_successors_map_.find(block);
if (asmi != augmented_successors_map_.end())
return &(*asmi).second;
auto smi = successors_map_.find(block);
return &(*smi).second;
};
}
LocalSingleStoreElimPass::GetBlocksFunction
LocalSingleStoreElimPass::AugmentedCFGPredecessorsFunction() const {
return [this](const ir::BasicBlock* block) {
auto apmi = augmented_predecessors_map_.find(block);
if (apmi != augmented_predecessors_map_.end())
return &(*apmi).second;
auto pmi = predecessors_map_.find(block);
return &(*pmi).second;
};
}
void LocalSingleStoreElimPass::CalculateImmediateDominators(
ir::Function* func) {
// Compute CFG
vector<ir::BasicBlock*> ordered_blocks;
predecessors_map_.clear();
successors_map_.clear();
for (auto& blk : *func) {
ordered_blocks.push_back(&blk);
blk.ForEachSuccessorLabel([&blk, &ordered_blocks, this](uint32_t sbid) {
successors_map_[&blk].push_back(label2block_[sbid]);
predecessors_map_[label2block_[sbid]].push_back(&blk);
});
}
// Compute Augmented CFG
augmented_successors_map_.clear();
augmented_predecessors_map_.clear();
successors_map_[&pseudo_exit_block_] = {};
predecessors_map_[&pseudo_entry_block_] = {};
auto succ_func = [this](const ir::BasicBlock* b)
{ return &successors_map_[b]; };
auto pred_func = [this](const ir::BasicBlock* b)
{ return &predecessors_map_[b]; };
CFA<ir::BasicBlock>::ComputeAugmentedCFG(
ordered_blocks,
&pseudo_entry_block_,
&pseudo_exit_block_,
&augmented_successors_map_,
&augmented_predecessors_map_,
succ_func,
pred_func);
// Compute Dominators
vector<const ir::BasicBlock*> postorder;
auto ignore_block = [](cbb_ptr) {};
auto ignore_edge = [](cbb_ptr, cbb_ptr) {};
spvtools::CFA<ir::BasicBlock>::DepthFirstTraversal(
ordered_blocks[0], AugmentedCFGSuccessorsFunction(),
ignore_block, [&](cbb_ptr b) { postorder.push_back(b); },
ignore_edge);
auto edges = spvtools::CFA<ir::BasicBlock>::CalculateDominators(
postorder, AugmentedCFGPredecessorsFunction());
idom_.clear();
for (auto edge : edges)
idom_[edge.first] = edge.second;
}
bool LocalSingleStoreElimPass::Dominates(
ir::BasicBlock* blk0, uint32_t idx0,
ir::BasicBlock* blk1, uint32_t idx1) {
if (blk0 == blk1)
return idx0 <= idx1;
ir::BasicBlock* b = blk1;
while (idom_[b] != b) {
b = idom_[b];
if (b == blk0)
return true;
}
return false;
}
bool LocalSingleStoreElimPass::SingleStoreProcess(ir::Function* func) {
CalculateImmediateDominators(func);
bool modified = false;
for (auto bi = func->begin(); bi != func->end(); ++bi) {
uint32_t instIdx = 0;
for (auto ii = bi->begin(); ii != bi->end(); ++ii, ++instIdx) {
if (ii->opcode() != SpvOpLoad)
continue;
uint32_t varId;
ir::Instruction* ptrInst = GetPtr(&*ii, &varId);
// Skip access chain loads
if (IsNonPtrAccessChain(ptrInst->opcode()))
continue;
if (ptrInst->opcode() != SpvOpVariable)
continue;
const auto vsi = ssa_var2store_.find(varId);
if (vsi == ssa_var2store_.end())
continue;
if (non_ssa_vars_.find(varId) != non_ssa_vars_.end())
continue;
// store must dominate load
if (!Dominates(store2blk_[vsi->second], store2idx_[vsi->second], &*bi, instIdx))
continue;
// Use store value as replacement id
uint32_t replId = vsi->second->GetSingleWordInOperand(kSpvStoreValId);
// replace all instances of the load's id with the SSA value's id
ReplaceAndDeleteLoad(&*ii, replId);
modified = true;
}
}
return modified;
}
bool LocalSingleStoreElimPass::HasLoads(uint32_t varId) const {
analysis::UseList* uses = def_use_mgr_->GetUses(varId);
if (uses == nullptr)
return false;
for (auto u : *uses) {
SpvOp op = u.inst->opcode();
// TODO(): The following is slightly conservative. Could be
// better handling of non-store/name.
if (IsNonPtrAccessChain(op) || op == SpvOpCopyObject) {
if (HasLoads(u.inst->result_id()))
return true;
}
else if (op != SpvOpStore && op != SpvOpName)
return true;
}
return false;
}
bool LocalSingleStoreElimPass::IsLiveVar(uint32_t varId) const {
// non-function scope vars are live
const ir::Instruction* varInst = def_use_mgr_->GetDef(varId);
assert(varInst->opcode() == SpvOpVariable);
const uint32_t varTypeId = varInst->type_id();
const ir::Instruction* varTypeInst = def_use_mgr_->GetDef(varTypeId);
if (varTypeInst->GetSingleWordInOperand(kSpvTypePointerStorageClass) !=
SpvStorageClassFunction)
return true;
// test if variable is loaded from
return HasLoads(varId);
}
bool LocalSingleStoreElimPass::IsLiveStore(ir::Instruction* storeInst) {
// get store's variable
uint32_t varId;
(void) GetPtr(storeInst, &varId);
return IsLiveVar(varId);
}
void LocalSingleStoreElimPass::AddStores(
uint32_t ptr_id, std::queue<ir::Instruction*>* insts) {
analysis::UseList* uses = def_use_mgr_->GetUses(ptr_id);
if (uses != nullptr) {
for (auto u : *uses) {
if (IsNonPtrAccessChain(u.inst->opcode()))
AddStores(u.inst->result_id(), insts);
else if (u.inst->opcode() == SpvOpStore)
insts->push(u.inst);
}
}
}
void LocalSingleStoreElimPass::DCEInst(ir::Instruction* inst) {
std::queue<ir::Instruction*> deadInsts;
deadInsts.push(inst);
while (!deadInsts.empty()) {
ir::Instruction* di = deadInsts.front();
// Don't delete labels
if (di->opcode() == SpvOpLabel) {
deadInsts.pop();
continue;
}
// Remember operands
std::queue<uint32_t> ids;
di->ForEachInId([&ids](uint32_t* iid) {
ids.push(*iid);
});
uint32_t varId = 0;
// Remember variable if dead load
if (di->opcode() == SpvOpLoad)
(void) GetPtr(di, &varId);
def_use_mgr_->KillInst(di);
// For all operands with no remaining uses, add their instruction
// to the dead instruction queue.
while (!ids.empty()) {
uint32_t id = ids.front();
analysis::UseList* uses = def_use_mgr_->GetUses(id);
if (uses == nullptr)
deadInsts.push(def_use_mgr_->GetDef(id));
ids.pop();
}
// if a load was deleted and it was the variable's
// last load, add all its stores to dead queue
if (varId != 0 && !IsLiveVar(varId))
AddStores(varId, &deadInsts);
deadInsts.pop();
}
}
bool LocalSingleStoreElimPass::SingleStoreDCE() {
bool modified = false;
for (auto v : ssa_var2store_) {
// check that it hasn't already been DCE'd
if (v.second->opcode() != SpvOpStore)
continue;
if (non_ssa_vars_.find(v.first) != non_ssa_vars_.end())
continue;
if (!IsLiveStore(v.second)) {
DCEInst(v.second);
modified = true;
}
}
return modified;
}
bool LocalSingleStoreElimPass::LocalSingleStoreElim(ir::Function* func) {
bool modified = false;
SingleStoreAnalyze(func);
if (ssa_var2store_.empty())
return false;
modified |= SingleStoreProcess(func);
modified |= SingleStoreDCE();
return modified;
}
void LocalSingleStoreElimPass::Initialize(ir::Module* module) {
module_ = module;
// Initialize function and block maps
id2function_.clear();
label2block_.clear();
for (auto& fn : *module_) {
id2function_[fn.result_id()] = &fn;
for (auto& blk : fn) {
uint32_t bid = blk.id();
label2block_[bid] = &blk;
}
}
// Initialize Target Type Caches
seen_target_vars_.clear();
seen_non_target_vars_.clear();
// Initialize Supported Ref Pointer Cache
supported_ref_ptrs_.clear();
// TODO: Reuse def/use (and other state) from previous passes
def_use_mgr_.reset(new analysis::DefUseManager(consumer(), module_));
// Initialize next unused Id
next_id_ = module_->id_bound();
};
Pass::Status LocalSingleStoreElimPass::ProcessImpl() {
// Assumes logical addressing only
if (module_->HasCapability(SpvCapabilityAddresses))
return Status::SuccessWithoutChange;
bool modified = false;
// Call Mem2Reg on all remaining functions.
for (auto& e : module_->entry_points()) {
ir::Function* fn =
id2function_[e.GetSingleWordOperand(kSpvEntryPointFunctionId)];
modified = modified || LocalSingleStoreElim(fn);
}
FinalizeNextId(module_);
return modified ? Status::SuccessWithChange : Status::SuccessWithoutChange;
}
LocalSingleStoreElimPass::LocalSingleStoreElimPass()
: module_(nullptr), def_use_mgr_(nullptr),
pseudo_entry_block_(std::unique_ptr<ir::Instruction>(
new ir::Instruction(SpvOpLabel, 0, 0, {}))),
pseudo_exit_block_(std::unique_ptr<ir::Instruction>(
new ir::Instruction(SpvOpLabel, 0, kInvalidId, {}))),
next_id_(0) {}
Pass::Status LocalSingleStoreElimPass::Process(ir::Module* module) {
Initialize(module);
return ProcessImpl();
}
} // namespace opt
} // namespace spvtools

220
source/opt/local_single_store_elim_pass.h Normal file
View File

@ -0,0 +1,220 @@
// Copyright (c) 2017 The Khronos Group Inc.
// Copyright (c) 2017 Valve Corporation
// Copyright (c) 2017 LunarG 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.
#ifndef LIBSPIRV_OPT_LOCAL_SINGLE_STORE_ELIM_PASS_H_
#define LIBSPIRV_OPT_LOCAL_SINGLE_STORE_ELIM_PASS_H_
#include <algorithm>
#include <map>
#include <queue>
#include <unordered_map>
#include <unordered_set>
#include <utility>
#include "basic_block.h"
#include "def_use_manager.h"
#include "module.h"
#include "pass.h"
namespace spvtools {
namespace opt {
// See optimizer.hpp for documentation.
class LocalSingleStoreElimPass : public Pass {
using cbb_ptr = const ir::BasicBlock*;
public:
LocalSingleStoreElimPass();
const char* name() const override { return "eliminate-local-single-store"; }
Status Process(ir::Module*) override;
private:
// Returns true if |opcode| is a non-ptr access chain op
bool IsNonPtrAccessChain(const SpvOp opcode) const;
// Returns true if |typeInst| is a scalar type
// or a vector or matrix
bool IsMathType(const ir::Instruction* typeInst) const;
// Returns true if |typeInst| is a math type or a struct or array
// of a math type.
bool IsTargetType(const ir::Instruction* typeInst) const;
// Given a load or store |ip|, return the pointer instruction.
// Also return the base variable's id in |varId|.
ir::Instruction* GetPtr(ir::Instruction* ip, uint32_t* varId);
// Return true if |varId| is a previously identified target variable.
// Return false if |varId| is a previously identified non-target variable.
// If variable is not cached, return true if variable is a function scope
// variable of target type, false otherwise. Updates caches of target
// and non-target variables.
bool IsTargetVar(uint32_t varId);
// Return true if all refs through |ptrId| are only loads or stores and
// cache ptrId in supported_ref_ptrs_.
bool HasOnlySupportedRefs(uint32_t ptrId);
// Find all function scope variables in |func| that are stored to
// only once (SSA) and map to their stored value id. Only analyze
// variables of scalar, vector, matrix types and struct and array
// types comprising only these types. Currently this analysis is
// is not done in the presence of function calls. TODO(): Allow
// analysis in the presence of function calls.
void SingleStoreAnalyze(ir::Function* func);
// Replace all instances of |loadInst|'s id with |replId| and delete
// |loadInst|.
void ReplaceAndDeleteLoad(ir::Instruction* loadInst, uint32_t replId);
using GetBlocksFunction =
std::function<const std::vector<ir::BasicBlock*>*(const ir::BasicBlock*)>;
/// Returns the block successors function for the augmented CFG.
GetBlocksFunction AugmentedCFGSuccessorsFunction() const;
/// Returns the block predecessors function for the augmented CFG.
GetBlocksFunction AugmentedCFGPredecessorsFunction() const;
// Calculate immediate dominators for |func|'s CFG. Leaves result
// in idom_. Entries for augmented CFG (pseudo blocks) are not created.
void CalculateImmediateDominators(ir::Function* func);
// Return true if instruction in |blk0| at ordinal position |idx0|
// dominates instruction in |blk1| at position |idx1|.
bool Dominates(ir::BasicBlock* blk0, uint32_t idx0,
ir::BasicBlock* blk1, uint32_t idx1);
// For each load of an SSA variable in |func|, replace all uses of
// the load with the value stored if the store dominates the load.
// Assumes that SingleStoreAnalyze() has just been run. Return true
// if any instructions are modified.
bool SingleStoreProcess(ir::Function* func);
// Return true if any instruction loads from |varId|
bool HasLoads(uint32_t varId) const;
// Return true if |varId| is not a function variable or if it has
// a load
bool IsLiveVar(uint32_t varId) const;
// Return true if |storeInst| is not a function variable or if its
// base variable has a load
bool IsLiveStore(ir::Instruction* storeInst);
// Add stores using |ptr_id| to |insts|
void AddStores(uint32_t ptr_id, std::queue<ir::Instruction*>* insts);
// Delete |inst| and iterate DCE on all its operands if they are now
// useless. If a load is deleted and its variable has no other loads,
// delete all its variable's stores.
void DCEInst(ir::Instruction* inst);
// Remove all stores to useless SSA variables. Remove useless
// access chains and variables as well. Assumes SingleStoreAnalyze
// and SingleStoreProcess has been run.
bool SingleStoreDCE();
// Do "single-store" optimization of function variables defined only
// with a single non-access-chain store in |func|. Replace all their
// non-access-chain loads with the value that is stored and eliminate
// any resulting dead code.
bool LocalSingleStoreElim(ir::Function* func);
// Save next available id into |module|.
inline void FinalizeNextId(ir::Module* module) {
module->SetIdBound(next_id_);
}
// Return next available id and generate next.
inline uint32_t TakeNextId() {
return next_id_++;
}
void Initialize(ir::Module* module);
Pass::Status ProcessImpl();
// Module this pass is processing
ir::Module* module_;
// Def-Uses for the module we are processing
std::unique_ptr<analysis::DefUseManager> def_use_mgr_;
// Map from function's result id to function
std::unordered_map<uint32_t, ir::Function*> id2function_;
// Map from block's label id to block
std::unordered_map<uint32_t, ir::BasicBlock*> label2block_;
// Map from SSA Variable to its single store
std::unordered_map<uint32_t, ir::Instruction*> ssa_var2store_;
// Map from store to its ordinal position in its block.
std::unordered_map<ir::Instruction*, uint32_t> store2idx_;
// Map from store to its block.
std::unordered_map<ir::Instruction*, ir::BasicBlock*> store2blk_;
// Set of non-SSA Variables
std::unordered_set<uint32_t> non_ssa_vars_;
// Cache of previously seen target types
std::unordered_set<uint32_t> seen_target_vars_;
// Cache of previously seen non-target types
std::unordered_set<uint32_t> seen_non_target_vars_;
// Variables with only supported references, ie. loads and stores using
// variable directly or through non-ptr access chains.
std::unordered_set<uint32_t> supported_ref_ptrs_;
// Augmented CFG Entry Block
ir::BasicBlock pseudo_entry_block_;
// Augmented CFG Exit Block
ir::BasicBlock pseudo_exit_block_;
// CFG Predecessors
std::unordered_map<const ir::BasicBlock*, std::vector<ir::BasicBlock*>>
predecessors_map_;
// CFG Successors
std::unordered_map<const ir::BasicBlock*, std::vector<ir::BasicBlock*>>
successors_map_;
// CFG Augmented Predecessors
std::unordered_map<const ir::BasicBlock*, std::vector<ir::BasicBlock*>>
augmented_predecessors_map_;
// CFG Augmented Successors
std::unordered_map<const ir::BasicBlock*, std::vector<ir::BasicBlock*>>
augmented_successors_map_;
// Immediate Dominator Map
// If block has no idom it points to itself.
std::unordered_map<ir::BasicBlock*, ir::BasicBlock*> idom_;
// Next unused ID
uint32_t next_id_;
};
} // namespace opt
} // namespace spvtools
#endif // LIBSPIRV_OPT_LOCAL_SINGLE_STORE_ELIM_PASS_H_

5
source/opt/optimizer.cpp
View File

@ -146,6 +146,11 @@ Optimizer::PassToken CreateLocalSingleBlockLoadStoreElimPass() {
MakeUnique<opt::LocalSingleBlockLoadStoreElimPass>());
}
Optimizer::PassToken CreateLocalSingleStoreElimPass() {
return MakeUnique<Optimizer::PassToken::Impl>(
MakeUnique<opt::LocalSingleStoreElimPass>());
}
Optimizer::PassToken CreateCompactIdsPass() {
return MakeUnique<Optimizer::PassToken::Impl>(
MakeUnique<opt::CompactIdsPass>());

1
source/opt/passes.h
View File

@ -23,6 +23,7 @@
#include "fold_spec_constant_op_and_composite_pass.h"
#include "inline_pass.h"
#include "local_single_block_elim_pass.h"
#include "local_single_store_elim_pass.h"
#include "freeze_spec_constant_value_pass.h"
#include "local_access_chain_convert_pass.h"
#include "null_pass.h"

5
test/opt/CMakeLists.txt
View File

@ -68,6 +68,11 @@ add_spvtools_unittest(TARGET pass_local_access_chain_convert
LIBS SPIRV-Tools-opt
)
add_spvtools_unittest(TARGET pass_local_single_store_elim
SRCS local_single_store_elim_test.cpp pass_utils.cpp
LIBS SPIRV-Tools-opt
)
add_spvtools_unittest(TARGET pass_eliminate_dead_const
SRCS eliminate_dead_const_test.cpp pass_utils.cpp
LIBS SPIRV-Tools-opt

521
test/opt/local_single_store_elim_test.cpp Normal file
View File

@ -0,0 +1,521 @@
// Copyright (c) 2017 Valve Corporation
// Copyright (c) 2017 LunarG 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 "pass_fixture.h"
#include "pass_utils.h"
namespace {
using namespace spvtools;
using LocalSingleStoreElimTest = PassTest<::testing::Test>;
TEST_F(LocalSingleStoreElimTest, PositiveAndNegative) {
// Single store to v is optimized. Multiple store to
// f is not optimized.
//
// #version 140
//
// in vec4 BaseColor;
// in float fi;
//
// void main()
// {
// vec4 v = BaseColor;
// float f = fi;
// if (f < 0)
// f = 0.0;
// gl_FragColor = v + f;
// }
const std::string predefs =
R"(OpCapability Shader
%1 = OpExtInstImport "GLSL.std.450"
OpMemoryModel Logical GLSL450
OpEntryPoint Fragment %main "main" %BaseColor %fi %gl_FragColor
OpExecutionMode %main OriginUpperLeft
OpSource GLSL 140
OpName %main "main"
OpName %v "v"
OpName %BaseColor "BaseColor"
OpName %f "f"
OpName %fi "fi"
OpName %gl_FragColor "gl_FragColor"
%void = OpTypeVoid
%9 = OpTypeFunction %void
%float = OpTypeFloat 32
%v4float = OpTypeVector %float 4
%_ptr_Function_v4float = OpTypePointer Function %v4float
%_ptr_Input_v4float = OpTypePointer Input %v4float
%BaseColor = OpVariable %_ptr_Input_v4float Input
%_ptr_Function_float = OpTypePointer Function %float
%_ptr_Input_float = OpTypePointer Input %float
%fi = OpVariable %_ptr_Input_float Input
%float_0 = OpConstant %float 0
%bool = OpTypeBool
%_ptr_Output_v4float = OpTypePointer Output %v4float
%gl_FragColor = OpVariable %_ptr_Output_v4float Output
)";
const std::string before =
R"(%main = OpFunction %void None %9
%19 = OpLabel
%v = OpVariable %_ptr_Function_v4float Function
%f = OpVariable %_ptr_Function_float Function
%20 = OpLoad %v4float %BaseColor
OpStore %v %20
%21 = OpLoad %float %fi
OpStore %f %21
%22 = OpLoad %float %f
%23 = OpFOrdLessThan %bool %22 %float_0
OpSelectionMerge %24 None
OpBranchConditional %23 %25 %24
%25 = OpLabel
OpStore %f %float_0
OpBranch %24
%24 = OpLabel
%26 = OpLoad %v4float %v
%27 = OpLoad %float %f
%28 = OpCompositeConstruct %v4float %27 %27 %27 %27
%29 = OpFAdd %v4float %26 %28
OpStore %gl_FragColor %29
OpReturn
OpFunctionEnd
)";
const std::string after =
R"(%main = OpFunction %void None %9
%19 = OpLabel
%v = OpVariable %_ptr_Function_v4float Function
%f = OpVariable %_ptr_Function_float Function
%20 = OpLoad %v4float %BaseColor
%21 = OpLoad %float %fi
OpStore %f %21
%22 = OpLoad %float %f
%23 = OpFOrdLessThan %bool %22 %float_0
OpSelectionMerge %24 None
OpBranchConditional %23 %25 %24
%25 = OpLabel
OpStore %f %float_0
OpBranch %24
%24 = OpLabel
%27 = OpLoad %float %f
%28 = OpCompositeConstruct %v4float %27 %27 %27 %27
%29 = OpFAdd %v4float %20 %28
OpStore %gl_FragColor %29
OpReturn
OpFunctionEnd
)";
SinglePassRunAndCheck<opt::LocalSingleStoreElimPass>(predefs + before,
predefs + after, true, true);
}
TEST_F(LocalSingleStoreElimTest, MultipleLoads) {
// Single store to multiple loads of v is optimized.
//
// #version 140
//
// in vec4 BaseColor;
// in float fi;
//
// void main()
// {
// vec4 v = BaseColor;
// float f = fi;
// if (f < 0)
// f = 0.0;
// gl_FragColor = v + f;
// }
const std::string predefs =
R"(OpCapability Shader
%1 = OpExtInstImport "GLSL.std.450"
OpMemoryModel Logical GLSL450
OpEntryPoint Fragment %main "main" %BaseColor %fi %gl_FragColor
OpExecutionMode %main OriginUpperLeft
OpSource GLSL 140
OpName %main "main"
OpName %v "v"
OpName %BaseColor "BaseColor"
OpName %fi "fi"
OpName %r "r"
OpName %gl_FragColor "gl_FragColor"
%void = OpTypeVoid
%9 = OpTypeFunction %void
%float = OpTypeFloat 32
%v4float = OpTypeVector %float 4
%_ptr_Function_v4float = OpTypePointer Function %v4float
%_ptr_Input_v4float = OpTypePointer Input %v4float
%BaseColor = OpVariable %_ptr_Input_v4float Input
%_ptr_Input_float = OpTypePointer Input %float
%fi = OpVariable %_ptr_Input_float Input
%float_0 = OpConstant %float 0
%bool = OpTypeBool
%float_1 = OpConstant %float 1
%_ptr_Output_v4float = OpTypePointer Output %v4float
%gl_FragColor = OpVariable %_ptr_Output_v4float Output
)";
const std::string before =
R"(%main = OpFunction %void None %9
%19 = OpLabel
%v = OpVariable %_ptr_Function_v4float Function
%r = OpVariable %_ptr_Function_v4float Function
%20 = OpLoad %v4float %BaseColor
OpStore %v %20
%21 = OpLoad %float %fi
%22 = OpFOrdLessThan %bool %21 %float_0
OpSelectionMerge %23 None
OpBranchConditional %22 %24 %25
%24 = OpLabel
%26 = OpLoad %v4float %v
OpStore %r %26
OpBranch %23
%25 = OpLabel
%27 = OpLoad %v4float %v
%28 = OpCompositeConstruct %v4float %float_1 %float_1 %float_1 %float_1
%29 = OpFSub %v4float %28 %27
OpStore %r %29
OpBranch %23
%23 = OpLabel
%30 = OpLoad %v4float %r
OpStore %gl_FragColor %30
OpReturn
OpFunctionEnd
)";
const std::string after =
R"(%main = OpFunction %void None %9
%19 = OpLabel
%v = OpVariable %_ptr_Function_v4float Function
%r = OpVariable %_ptr_Function_v4float Function
%20 = OpLoad %v4float %BaseColor
%21 = OpLoad %float %fi
%22 = OpFOrdLessThan %bool %21 %float_0
OpSelectionMerge %23 None
OpBranchConditional %22 %24 %25
%24 = OpLabel
OpStore %r %20
OpBranch %23
%25 = OpLabel
%28 = OpCompositeConstruct %v4float %float_1 %float_1 %float_1 %float_1
%29 = OpFSub %v4float %28 %20
OpStore %r %29
OpBranch %23
%23 = OpLabel
%30 = OpLoad %v4float %r
OpStore %gl_FragColor %30
OpReturn
OpFunctionEnd
)";
SinglePassRunAndCheck<opt::LocalSingleStoreElimPass>(predefs + before,
predefs + after, true, true);
}
TEST_F(LocalSingleStoreElimTest, NoStoreElimWithInterveningAccessChainLoad) {
// Last load of v is eliminated, but access chain load and store of v isn't
//
// #version 140
//
// in vec4 BaseColor;
//
// void main()
// {
// vec4 v = BaseColor;
// float f = v[3];
// gl_FragColor = v * f;
// }
const std::string predefs =
R"(OpCapability Shader
%1 = OpExtInstImport "GLSL.std.450"
OpMemoryModel Logical GLSL450
OpEntryPoint Fragment %main "main" %BaseColor %gl_FragColor
OpExecutionMode %main OriginUpperLeft
OpSource GLSL 140
OpName %main "main"
OpName %v "v"
OpName %BaseColor "BaseColor"
OpName %f "f"
OpName %gl_FragColor "gl_FragColor"
%void = OpTypeVoid
%8 = OpTypeFunction %void
%float = OpTypeFloat 32
%v4float = OpTypeVector %float 4
%_ptr_Function_v4float = OpTypePointer Function %v4float
%_ptr_Input_v4float = OpTypePointer Input %v4float
%BaseColor = OpVariable %_ptr_Input_v4float Input
%_ptr_Function_float = OpTypePointer Function %float
%uint = OpTypeInt 32 0
%uint_3 = OpConstant %uint 3
%_ptr_Output_v4float = OpTypePointer Output %v4float
%gl_FragColor = OpVariable %_ptr_Output_v4float Output
)";
const std::string before =
R"(%main = OpFunction %void None %8
%17 = OpLabel
%v = OpVariable %_ptr_Function_v4float Function
%f = OpVariable %_ptr_Function_float Function
%18 = OpLoad %v4float %BaseColor
OpStore %v %18
%19 = OpAccessChain %_ptr_Function_float %v %uint_3
%20 = OpLoad %float %19
OpStore %f %20
%21 = OpLoad %v4float %v
%22 = OpLoad %float %f
%23 = OpVectorTimesScalar %v4float %21 %22
OpStore %gl_FragColor %23
OpReturn
OpFunctionEnd
)";
const std::string after =
R"(%main = OpFunction %void None %8
%17 = OpLabel
%v = OpVariable %_ptr_Function_v4float Function
%f = OpVariable %_ptr_Function_float Function
%18 = OpLoad %v4float %BaseColor
OpStore %v %18
%19 = OpAccessChain %_ptr_Function_float %v %uint_3
%20 = OpLoad %float %19
%23 = OpVectorTimesScalar %v4float %18 %20
OpStore %gl_FragColor %23
OpReturn
OpFunctionEnd
)";
SinglePassRunAndCheck<opt::LocalSingleStoreElimPass>(predefs + before,
predefs + after, true, true);
}
TEST_F(LocalSingleStoreElimTest, NoReplaceOfDominatingPartialStore) {
// Note: SPIR-V hand edited to initialize v to vec4(0.0)
//
// #version 140
//
// in vec4 BaseColor;
//
// void main()
// {
// vec4 v;
// float v[1] = 1.0;
// gl_FragColor = v;
// }
const std::string assembly =
R"(OpCapability Shader
%1 = OpExtInstImport "GLSL.std.450"
OpMemoryModel Logical GLSL450
OpEntryPoint Fragment %main "main" %gl_FragColor %BaseColor
OpExecutionMode %main OriginUpperLeft
OpSource GLSL 140
OpName %main "main"
OpName %v "v"
OpName %gl_FragColor "gl_FragColor"
OpName %BaseColor "BaseColor"
%void = OpTypeVoid
%7 = OpTypeFunction %void
%float = OpTypeFloat 32
%v4float = OpTypeVector %float 4
%_ptr_Function_v4float = OpTypePointer Function %v4float
%float_0 = OpConstant %float 0
%12 = OpConstantComposite %v4float %float_0 %float_0 %float_0 %float_0
%float_1 = OpConstant %float 1
%uint = OpTypeInt 32 0
%uint_1 = OpConstant %uint 1
%_ptr_Function_float = OpTypePointer Function %float
%_ptr_Output_v4float = OpTypePointer Output %v4float
%gl_FragColor = OpVariable %_ptr_Output_v4float Output
%_ptr_Input_v4float = OpTypePointer Input %v4float
%BaseColor = OpVariable %_ptr_Input_v4float Input
%main = OpFunction %void None %7
%19 = OpLabel
%v = OpVariable %_ptr_Function_v4float Function %12
%20 = OpAccessChain %_ptr_Function_float %v %uint_1
OpStore %20 %float_1
%21 = OpLoad %v4float %v
OpStore %gl_FragColor %21
OpReturn
OpFunctionEnd
)";
SinglePassRunAndCheck<opt::LocalSingleStoreElimPass>(
assembly, assembly, true, true);
}
TEST_F(LocalSingleStoreElimTest, NoReplaceInPresenceOfUnsupportedInst) {
// Note: PositiveNegative test hand edited to insert OpCopyObject
//
// #version 140
//
// in vec4 BaseColor;
// in float fi;
//
// void main()
// {
// vec4 v = BaseColor;
// float f = fi;
// if (f < 0)
// f = 0.0;
// gl_FragColor = v + f;
// }
const std::string assembly =
R"(OpCapability Shader
%1 = OpExtInstImport "GLSL.std.450"
OpMemoryModel Logical GLSL450
OpEntryPoint Fragment %main "main" %BaseColor %fi %gl_FragColor
OpExecutionMode %main OriginUpperLeft
OpSource GLSL 140
OpName %main "main"
OpName %v "v"
OpName %BaseColor "BaseColor"
OpName %f "f"
OpName %fi "fi"
OpName %gl_FragColor "gl_FragColor"
%void = OpTypeVoid
%9 = OpTypeFunction %void
%float = OpTypeFloat 32
%v4float = OpTypeVector %float 4
%_ptr_Function_v4float = OpTypePointer Function %v4float
%_ptr_Input_v4float = OpTypePointer Input %v4float
%BaseColor = OpVariable %_ptr_Input_v4float Input
%_ptr_Function_float = OpTypePointer Function %float
%_ptr_Input_float = OpTypePointer Input %float
%fi = OpVariable %_ptr_Input_float Input
%float_0 = OpConstant %float 0
%bool = OpTypeBool
%_ptr_Output_v4float = OpTypePointer Output %v4float
%gl_FragColor = OpVariable %_ptr_Output_v4float Output
%main = OpFunction %void None %9
%19 = OpLabel
%v = OpVariable %_ptr_Function_v4float Function
%f = OpVariable %_ptr_Function_float Function
%20 = OpLoad %v4float %BaseColor
OpStore %v %20
%21 = OpLoad %float %fi
OpStore %f %21
%22 = OpLoad %float %f
%23 = OpFOrdLessThan %bool %22 %float_0
OpSelectionMerge %24 None
OpBranchConditional %23 %25 %24
%25 = OpLabel
OpStore %f %float_0
OpBranch %24
%24 = OpLabel
%26 = OpCopyObject %_ptr_Function_v4float %v
%27 = OpLoad %v4float %26
%28 = OpLoad %float %f
%29 = OpCompositeConstruct %v4float %28 %28 %28 %28
%30 = OpFAdd %v4float %27 %29
OpStore %gl_FragColor %30
OpReturn
OpFunctionEnd
)";
SinglePassRunAndCheck<opt::LocalSingleStoreElimPass>(
assembly, assembly, true, true);
}
TEST_F(LocalSingleStoreElimTest, NoOptIfStoreNotDominating) {
// Single store to f not optimized because it does not dominate
// the load.
//
// #version 140
//
// in vec4 BaseColor;
// in float fi;
//
// void main()
// {
// float f;
// if (fi < 0)
// f = 0.5;
// if (fi < 0)
// gl_FragColor = BaseColor * f;
// else
// gl_FragColor = BaseColor;
// }
const std::string assembly =
R"(OpCapability Shader
%1 = OpExtInstImport "GLSL.std.450"
OpMemoryModel Logical GLSL450
OpEntryPoint Fragment %main "main" %fi %gl_FragColor %BaseColor
OpExecutionMode %main OriginUpperLeft
OpSource GLSL 140
OpName %main "main"
OpName %fi "fi"
OpName %f "f"
OpName %gl_FragColor "gl_FragColor"
OpName %BaseColor "BaseColor"
%void = OpTypeVoid
%8 = OpTypeFunction %void
%float = OpTypeFloat 32
%_ptr_Input_float = OpTypePointer Input %float
%fi = OpVariable %_ptr_Input_float Input
%float_0 = OpConstant %float 0
%bool = OpTypeBool
%_ptr_Function_float = OpTypePointer Function %float
%float_0_5 = OpConstant %float 0.5
%v4float = OpTypeVector %float 4
%_ptr_Output_v4float = OpTypePointer Output %v4float
%gl_FragColor = OpVariable %_ptr_Output_v4float Output
%_ptr_Input_v4float = OpTypePointer Input %v4float
%BaseColor = OpVariable %_ptr_Input_v4float Input
%main = OpFunction %void None %8
%18 = OpLabel
%f = OpVariable %_ptr_Function_float Function
%19 = OpLoad %float %fi
%20 = OpFOrdLessThan %bool %19 %float_0
OpSelectionMerge %21 None
OpBranchConditional %20 %22 %21
%22 = OpLabel
OpStore %f %float_0_5
OpBranch %21
%21 = OpLabel
%23 = OpLoad %float %fi
%24 = OpFOrdLessThan %bool %23 %float_0
OpSelectionMerge %25 None
OpBranchConditional %24 %26 %27
%26 = OpLabel
%28 = OpLoad %v4float %BaseColor
%29 = OpLoad %float %f
%30 = OpVectorTimesScalar %v4float %28 %29
OpStore %gl_FragColor %30
OpBranch %25
%27 = OpLabel
%31 = OpLoad %v4float %BaseColor
OpStore %gl_FragColor %31
OpBranch %25
%25 = OpLabel
OpReturn
OpFunctionEnd
)";
SinglePassRunAndCheck<opt::LocalSingleStoreElimPass>(assembly, assembly,
true, true);
}
// TODO(greg-lunarg): Add tests to verify handling of these cases:
//
// Other types
// Others?
} // anonymous namespace

2
tools/opt/opt.cpp
View File

@ -137,6 +137,8 @@ int main(int argc, char** argv) {
optimizer.RegisterPass(CreateLocalAccessChainConvertPass());
} else if (0 == strcmp(cur_arg, "--eliminate-local-single-block")) {
optimizer.RegisterPass(CreateLocalSingleBlockLoadStoreElimPass());
} else if (0 == strcmp(cur_arg, "--eliminate-local-single-store")) {
optimizer.RegisterPass(CreateLocalSingleStoreElimPass());
} else if (0 == strcmp(cur_arg, "--eliminate-dead-const")) {
optimizer.RegisterPass(CreateEliminateDeadConstantPass());
} else if (0 == strcmp(cur_arg, "--fold-spec-const-op-composite")) {