SPIRV-Tools/source/opt/aggressive_dead_code_elim_pass.cpp

940 lines
35 KiB
C++
Raw Normal View History

// Copyright (c) 2017 The Khronos Group Inc.
// Copyright (c) 2017 Valve Corporation
// Copyright (c) 2017 LunarG Inc.
// 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 "source/opt/aggressive_dead_code_elim_pass.h"
#include <memory>
#include <stack>
#include "source/cfa.h"
#include "source/latest_version_glsl_std_450_header.h"
#include "source/opt/iterator.h"
#include "source/opt/reflect.h"
#include "source/spirv_constant.h"
namespace spvtools {
namespace opt {
namespace {
const uint32_t kTypePointerStorageClassInIdx = 0;
const uint32_t kEntryPointFunctionIdInIdx = 1;
const uint32_t kSelectionMergeMergeBlockIdInIdx = 0;
const uint32_t kLoopMergeMergeBlockIdInIdx = 0;
const uint32_t kLoopMergeContinueBlockIdInIdx = 1;
const uint32_t kCopyMemoryTargetAddrInIdx = 0;
const uint32_t kCopyMemorySourceAddrInIdx = 1;
// Sorting functor to present annotation instructions in an easy-to-process
// order. The functor orders by opcode first and falls back on unique id
// ordering if both instructions have the same opcode.
//
// Desired priority:
// SpvOpGroupDecorate
// SpvOpGroupMemberDecorate
// SpvOpDecorate
// SpvOpMemberDecorate
// SpvOpDecorateId
// SpvOpDecorateStringGOOGLE
// SpvOpDecorationGroup
struct DecorationLess {
bool operator()(const Instruction* lhs, const Instruction* rhs) const {
assert(lhs && rhs);
SpvOp lhsOp = lhs->opcode();
SpvOp rhsOp = rhs->opcode();
if (lhsOp != rhsOp) {
#define PRIORITY_CASE(opcode) \
if (lhsOp == opcode && rhsOp != opcode) return true; \
if (rhsOp == opcode && lhsOp != opcode) return false;
// OpGroupDecorate and OpGroupMember decorate are highest priority to
// eliminate dead targets early and simplify subsequent checks.
PRIORITY_CASE(SpvOpGroupDecorate)
PRIORITY_CASE(SpvOpGroupMemberDecorate)
PRIORITY_CASE(SpvOpDecorate)
PRIORITY_CASE(SpvOpMemberDecorate)
PRIORITY_CASE(SpvOpDecorateId)
PRIORITY_CASE(SpvOpDecorateStringGOOGLE)
// OpDecorationGroup is lowest priority to ensure use/def chains remain
// usable for instructions that target this group.
PRIORITY_CASE(SpvOpDecorationGroup)
#undef PRIORITY_CASE
}
// Fall back to maintain total ordering (compare unique ids).
return *lhs < *rhs;
}
};
} // namespace
bool AggressiveDCEPass::IsVarOfStorage(uint32_t varId, uint32_t storageClass) {
if (varId == 0) return false;
const Instruction* varInst = get_def_use_mgr()->GetDef(varId);
const SpvOp op = varInst->opcode();
if (op != SpvOpVariable) return false;
const uint32_t varTypeId = varInst->type_id();
const Instruction* varTypeInst = get_def_use_mgr()->GetDef(varTypeId);
if (varTypeInst->opcode() != SpvOpTypePointer) return false;
return varTypeInst->GetSingleWordInOperand(kTypePointerStorageClassInIdx) ==
storageClass;
}
bool AggressiveDCEPass::IsLocalVar(uint32_t varId) {
if (IsVarOfStorage(varId, SpvStorageClassFunction)) {
return true;
}
if (!private_like_local_) {
return false;
}
return IsVarOfStorage(varId, SpvStorageClassPrivate) ||
IsVarOfStorage(varId, SpvStorageClassWorkgroup);
}
void AggressiveDCEPass::AddStores(uint32_t ptrId) {
get_def_use_mgr()->ForEachUser(ptrId, [this, ptrId](Instruction* user) {
Adding new def -> use mapping container Replaced representation of uses * Changed uses from unordered_map<uint32_t, UseList> to set<pairInstruction*, Instruction*>> * Replaced GetUses with ForEachUser and ForEachUse functions * updated passes to use new functions * partially updated tests * lots of cleanup still todo Adding an unique id to Instruction generated by IRContext Each instruction is given an unique id that can be used for ordering purposes. The ids are generated via the IRContext. Major changes: * Instructions now contain a uint32_t for unique id and a cached context pointer * Most constructors have been modified to take a context as input * unfortunately I cannot remove the default and copy constructors, but developers should avoid these * Added accessors to parents of basic block and function * Removed the copy constructors for BasicBlock and Function and replaced them with Clone functions * Reworked BuildModule to return an IRContext owning the built module * Since all instructions require a context, the context now becomes the basic unit for IR * Added a constructor to context to create an owned module internally * Replaced uses of Instruction's copy constructor with Clone whereever I found them * Reworked the linker functionality to perform clones into a different context instead of moves * Updated many tests to be consistent with the above changes * Still need to add new tests to cover added functionality * Added comparison operators to Instruction Adding tests for Instruction, IRContext and IR loading Fixed some header comments for BuildModule Fixes to get tests passing again * Reordered two linker steps to avoid use/def problems * Fixed def/use manager uses in merge return pass * Added early return for GetAnnotations * Changed uses of Instruction::ToNop in passes to IRContext::KillInst Simplifying the uses for some contexts in passes
2017-11-14 19:11:50 +00:00
switch (user->opcode()) {
case SpvOpAccessChain:
case SpvOpInBoundsAccessChain:
Adding new def -> use mapping container Replaced representation of uses * Changed uses from unordered_map<uint32_t, UseList> to set<pairInstruction*, Instruction*>> * Replaced GetUses with ForEachUser and ForEachUse functions * updated passes to use new functions * partially updated tests * lots of cleanup still todo Adding an unique id to Instruction generated by IRContext Each instruction is given an unique id that can be used for ordering purposes. The ids are generated via the IRContext. Major changes: * Instructions now contain a uint32_t for unique id and a cached context pointer * Most constructors have been modified to take a context as input * unfortunately I cannot remove the default and copy constructors, but developers should avoid these * Added accessors to parents of basic block and function * Removed the copy constructors for BasicBlock and Function and replaced them with Clone functions * Reworked BuildModule to return an IRContext owning the built module * Since all instructions require a context, the context now becomes the basic unit for IR * Added a constructor to context to create an owned module internally * Replaced uses of Instruction's copy constructor with Clone whereever I found them * Reworked the linker functionality to perform clones into a different context instead of moves * Updated many tests to be consistent with the above changes * Still need to add new tests to cover added functionality * Added comparison operators to Instruction Adding tests for Instruction, IRContext and IR loading Fixed some header comments for BuildModule Fixes to get tests passing again * Reordered two linker steps to avoid use/def problems * Fixed def/use manager uses in merge return pass * Added early return for GetAnnotations * Changed uses of Instruction::ToNop in passes to IRContext::KillInst Simplifying the uses for some contexts in passes
2017-11-14 19:11:50 +00:00
case SpvOpCopyObject:
this->AddStores(user->result_id());
break;
case SpvOpLoad:
break;
case SpvOpCopyMemory:
case SpvOpCopyMemorySized:
if (user->GetSingleWordInOperand(kCopyMemoryTargetAddrInIdx) == ptrId) {
AddToWorklist(user);
}
break;
Adding new def -> use mapping container Replaced representation of uses * Changed uses from unordered_map<uint32_t, UseList> to set<pairInstruction*, Instruction*>> * Replaced GetUses with ForEachUser and ForEachUse functions * updated passes to use new functions * partially updated tests * lots of cleanup still todo Adding an unique id to Instruction generated by IRContext Each instruction is given an unique id that can be used for ordering purposes. The ids are generated via the IRContext. Major changes: * Instructions now contain a uint32_t for unique id and a cached context pointer * Most constructors have been modified to take a context as input * unfortunately I cannot remove the default and copy constructors, but developers should avoid these * Added accessors to parents of basic block and function * Removed the copy constructors for BasicBlock and Function and replaced them with Clone functions * Reworked BuildModule to return an IRContext owning the built module * Since all instructions require a context, the context now becomes the basic unit for IR * Added a constructor to context to create an owned module internally * Replaced uses of Instruction's copy constructor with Clone whereever I found them * Reworked the linker functionality to perform clones into a different context instead of moves * Updated many tests to be consistent with the above changes * Still need to add new tests to cover added functionality * Added comparison operators to Instruction Adding tests for Instruction, IRContext and IR loading Fixed some header comments for BuildModule Fixes to get tests passing again * Reordered two linker steps to avoid use/def problems * Fixed def/use manager uses in merge return pass * Added early return for GetAnnotations * Changed uses of Instruction::ToNop in passes to IRContext::KillInst Simplifying the uses for some contexts in passes
2017-11-14 19:11:50 +00:00
// If default, assume it stores e.g. frexp, modf, function call
case SpvOpStore:
Adding new def -> use mapping container Replaced representation of uses * Changed uses from unordered_map<uint32_t, UseList> to set<pairInstruction*, Instruction*>> * Replaced GetUses with ForEachUser and ForEachUse functions * updated passes to use new functions * partially updated tests * lots of cleanup still todo Adding an unique id to Instruction generated by IRContext Each instruction is given an unique id that can be used for ordering purposes. The ids are generated via the IRContext. Major changes: * Instructions now contain a uint32_t for unique id and a cached context pointer * Most constructors have been modified to take a context as input * unfortunately I cannot remove the default and copy constructors, but developers should avoid these * Added accessors to parents of basic block and function * Removed the copy constructors for BasicBlock and Function and replaced them with Clone functions * Reworked BuildModule to return an IRContext owning the built module * Since all instructions require a context, the context now becomes the basic unit for IR * Added a constructor to context to create an owned module internally * Replaced uses of Instruction's copy constructor with Clone whereever I found them * Reworked the linker functionality to perform clones into a different context instead of moves * Updated many tests to be consistent with the above changes * Still need to add new tests to cover added functionality * Added comparison operators to Instruction Adding tests for Instruction, IRContext and IR loading Fixed some header comments for BuildModule Fixes to get tests passing again * Reordered two linker steps to avoid use/def problems * Fixed def/use manager uses in merge return pass * Added early return for GetAnnotations * Changed uses of Instruction::ToNop in passes to IRContext::KillInst Simplifying the uses for some contexts in passes
2017-11-14 19:11:50 +00:00
default:
AddToWorklist(user);
Adding new def -> use mapping container Replaced representation of uses * Changed uses from unordered_map<uint32_t, UseList> to set<pairInstruction*, Instruction*>> * Replaced GetUses with ForEachUser and ForEachUse functions * updated passes to use new functions * partially updated tests * lots of cleanup still todo Adding an unique id to Instruction generated by IRContext Each instruction is given an unique id that can be used for ordering purposes. The ids are generated via the IRContext. Major changes: * Instructions now contain a uint32_t for unique id and a cached context pointer * Most constructors have been modified to take a context as input * unfortunately I cannot remove the default and copy constructors, but developers should avoid these * Added accessors to parents of basic block and function * Removed the copy constructors for BasicBlock and Function and replaced them with Clone functions * Reworked BuildModule to return an IRContext owning the built module * Since all instructions require a context, the context now becomes the basic unit for IR * Added a constructor to context to create an owned module internally * Replaced uses of Instruction's copy constructor with Clone whereever I found them * Reworked the linker functionality to perform clones into a different context instead of moves * Updated many tests to be consistent with the above changes * Still need to add new tests to cover added functionality * Added comparison operators to Instruction Adding tests for Instruction, IRContext and IR loading Fixed some header comments for BuildModule Fixes to get tests passing again * Reordered two linker steps to avoid use/def problems * Fixed def/use manager uses in merge return pass * Added early return for GetAnnotations * Changed uses of Instruction::ToNop in passes to IRContext::KillInst Simplifying the uses for some contexts in passes
2017-11-14 19:11:50 +00:00
break;
}
Adding new def -> use mapping container Replaced representation of uses * Changed uses from unordered_map<uint32_t, UseList> to set<pairInstruction*, Instruction*>> * Replaced GetUses with ForEachUser and ForEachUse functions * updated passes to use new functions * partially updated tests * lots of cleanup still todo Adding an unique id to Instruction generated by IRContext Each instruction is given an unique id that can be used for ordering purposes. The ids are generated via the IRContext. Major changes: * Instructions now contain a uint32_t for unique id and a cached context pointer * Most constructors have been modified to take a context as input * unfortunately I cannot remove the default and copy constructors, but developers should avoid these * Added accessors to parents of basic block and function * Removed the copy constructors for BasicBlock and Function and replaced them with Clone functions * Reworked BuildModule to return an IRContext owning the built module * Since all instructions require a context, the context now becomes the basic unit for IR * Added a constructor to context to create an owned module internally * Replaced uses of Instruction's copy constructor with Clone whereever I found them * Reworked the linker functionality to perform clones into a different context instead of moves * Updated many tests to be consistent with the above changes * Still need to add new tests to cover added functionality * Added comparison operators to Instruction Adding tests for Instruction, IRContext and IR loading Fixed some header comments for BuildModule Fixes to get tests passing again * Reordered two linker steps to avoid use/def problems * Fixed def/use manager uses in merge return pass * Added early return for GetAnnotations * Changed uses of Instruction::ToNop in passes to IRContext::KillInst Simplifying the uses for some contexts in passes
2017-11-14 19:11:50 +00:00
});
}
bool AggressiveDCEPass::AllExtensionsSupported() const {
// If any extension not in whitelist, return false
for (auto& ei : get_module()->extensions()) {
const char* extName =
reinterpret_cast<const char*>(&ei.GetInOperand(0).words[0]);
if (extensions_whitelist_.find(extName) == extensions_whitelist_.end())
return false;
}
return true;
}
bool AggressiveDCEPass::IsDead(Instruction* inst) {
if (IsLive(inst)) return false;
if ((inst->IsBranch() || inst->opcode() == SpvOpUnreachable) &&
!IsStructuredHeader(context()->get_instr_block(inst), nullptr, nullptr,
nullptr))
return false;
return true;
}
bool AggressiveDCEPass::IsTargetDead(Instruction* inst) {
const uint32_t tId = inst->GetSingleWordInOperand(0);
Instruction* tInst = get_def_use_mgr()->GetDef(tId);
if (IsAnnotationInst(tInst->opcode())) {
// This must be a decoration group. We go through annotations in a specific
// order. So if this is not used by any group or group member decorates, it
// is dead.
assert(tInst->opcode() == SpvOpDecorationGroup);
bool dead = true;
get_def_use_mgr()->ForEachUser(tInst, [&dead](Instruction* user) {
if (user->opcode() == SpvOpGroupDecorate ||
user->opcode() == SpvOpGroupMemberDecorate)
dead = false;
});
return dead;
}
return IsDead(tInst);
}
void AggressiveDCEPass::ProcessLoad(uint32_t varId) {
// Only process locals
if (!IsLocalVar(varId)) return;
// Return if already processed
if (live_local_vars_.find(varId) != live_local_vars_.end()) return;
// Mark all stores to varId as live
AddStores(varId);
// Cache varId as processed
live_local_vars_.insert(varId);
}
bool AggressiveDCEPass::IsStructuredHeader(BasicBlock* bp,
Instruction** mergeInst,
Instruction** branchInst,
uint32_t* mergeBlockId) {
if (!bp) return false;
Instruction* mi = bp->GetMergeInst();
if (mi == nullptr) return false;
Instruction* bri = &*bp->tail();
if (branchInst != nullptr) *branchInst = bri;
if (mergeInst != nullptr) *mergeInst = mi;
if (mergeBlockId != nullptr) *mergeBlockId = mi->GetSingleWordInOperand(0);
return true;
}
void AggressiveDCEPass::ComputeBlock2HeaderMaps(
std::list<BasicBlock*>& structuredOrder) {
block2headerBranch_.clear();
header2nextHeaderBranch_.clear();
branch2merge_.clear();
structured_order_index_.clear();
std::stack<Instruction*> currentHeaderBranch;
currentHeaderBranch.push(nullptr);
uint32_t currentMergeBlockId = 0;
uint32_t index = 0;
for (auto bi = structuredOrder.begin(); bi != structuredOrder.end();
++bi, ++index) {
structured_order_index_[*bi] = index;
// If this block is the merge block of the current control construct,
// we are leaving the current construct so we must update state
if ((*bi)->id() == currentMergeBlockId) {
currentHeaderBranch.pop();
Instruction* chb = currentHeaderBranch.top();
if (chb != nullptr)
currentMergeBlockId = branch2merge_[chb]->GetSingleWordInOperand(0);
}
Instruction* mergeInst;
Instruction* branchInst;
uint32_t mergeBlockId;
bool is_header =
IsStructuredHeader(*bi, &mergeInst, &branchInst, &mergeBlockId);
// Map header block to next enclosing header.
if (is_header) header2nextHeaderBranch_[*bi] = currentHeaderBranch.top();
// If this is a loop header, update state first so the block will map to
// itself.
if (is_header && mergeInst->opcode() == SpvOpLoopMerge) {
currentHeaderBranch.push(branchInst);
branch2merge_[branchInst] = mergeInst;
currentMergeBlockId = mergeBlockId;
}
// Map the block to the current construct.
block2headerBranch_[*bi] = currentHeaderBranch.top();
// If this is an if header, update state so following blocks map to the if.
if (is_header && mergeInst->opcode() == SpvOpSelectionMerge) {
currentHeaderBranch.push(branchInst);
branch2merge_[branchInst] = mergeInst;
currentMergeBlockId = mergeBlockId;
}
}
}
void AggressiveDCEPass::AddBranch(uint32_t labelId, BasicBlock* bp) {
std::unique_ptr<Instruction> newBranch(
new Instruction(context(), SpvOpBranch, 0, 0,
{{spv_operand_type_t::SPV_OPERAND_TYPE_ID, {labelId}}}));
context()->AnalyzeDefUse(&*newBranch);
context()->set_instr_block(&*newBranch, bp);
bp->AddInstruction(std::move(newBranch));
}
void AggressiveDCEPass::AddBreaksAndContinuesToWorklist(
Instruction* mergeInst) {
assert(mergeInst->opcode() == SpvOpSelectionMerge ||
mergeInst->opcode() == SpvOpLoopMerge);
BasicBlock* header = context()->get_instr_block(mergeInst);
uint32_t headerIndex = structured_order_index_[header];
const uint32_t mergeId = mergeInst->GetSingleWordInOperand(0);
BasicBlock* merge = context()->get_instr_block(mergeId);
uint32_t mergeIndex = structured_order_index_[merge];
get_def_use_mgr()->ForEachUser(
mergeId, [headerIndex, mergeIndex, this](Instruction* user) {
if (!user->IsBranch()) return;
BasicBlock* block = context()->get_instr_block(user);
uint32_t index = structured_order_index_[block];
if (headerIndex < index && index < mergeIndex) {
// This is a break from the loop.
AddToWorklist(user);
// Add branch's merge if there is one.
Instruction* userMerge = branch2merge_[user];
if (userMerge != nullptr) AddToWorklist(userMerge);
}
});
if (mergeInst->opcode() != SpvOpLoopMerge) {
return;
}
// For loops we need to find the continues as well.
const uint32_t contId =
mergeInst->GetSingleWordInOperand(kLoopMergeContinueBlockIdInIdx);
get_def_use_mgr()->ForEachUser(contId, [&contId, this](Instruction* user) {
SpvOp op = user->opcode();
if (op == SpvOpBranchConditional || op == SpvOpSwitch) {
// A conditional branch or switch can only be a continue if it does not
// have a merge instruction or its merge block is not the continue block.
Instruction* hdrMerge = branch2merge_[user];
if (hdrMerge != nullptr && hdrMerge->opcode() == SpvOpSelectionMerge) {
uint32_t hdrMergeId =
hdrMerge->GetSingleWordInOperand(kSelectionMergeMergeBlockIdInIdx);
if (hdrMergeId == contId) return;
// Need to mark merge instruction too
AddToWorklist(hdrMerge);
}
} else if (op == SpvOpBranch) {
// An unconditional branch can only be a continue if it is not
// branching to its own merge block.
BasicBlock* blk = context()->get_instr_block(user);
Instruction* hdrBranch = block2headerBranch_[blk];
if (hdrBranch == nullptr) return;
Instruction* hdrMerge = branch2merge_[hdrBranch];
if (hdrMerge->opcode() == SpvOpLoopMerge) return;
uint32_t hdrMergeId =
hdrMerge->GetSingleWordInOperand(kSelectionMergeMergeBlockIdInIdx);
if (contId == hdrMergeId) return;
} else {
return;
}
AddToWorklist(user);
});
}
bool AggressiveDCEPass::AggressiveDCE(Function* func) {
// Mark function parameters as live.
AddToWorklist(&func->DefInst());
func->ForEachParam(
[this](const Instruction* param) {
AddToWorklist(const_cast<Instruction*>(param));
},
false);
// Compute map from block to controlling conditional branch
std::list<BasicBlock*> structuredOrder;
cfg()->ComputeStructuredOrder(func, &*func->begin(), &structuredOrder);
ComputeBlock2HeaderMaps(structuredOrder);
bool modified = false;
// Add instructions with external side effects to worklist. Also add branches
// EXCEPT those immediately contained in an "if" selection construct or a loop
// or continue construct.
// TODO(greg-lunarg): Handle Frexp, Modf more optimally
call_in_func_ = false;
func_is_entry_point_ = false;
private_stores_.clear();
// Stacks to keep track of when we are inside an if- or loop-construct.
// When immediately inside an if- or loop-construct, we do not initially
// mark branches live. All other branches must be marked live.
std::stack<bool> assume_branches_live;
std::stack<uint32_t> currentMergeBlockId;
// Push sentinel values on stack for when outside of any control flow.
assume_branches_live.push(true);
currentMergeBlockId.push(0);
for (auto bi = structuredOrder.begin(); bi != structuredOrder.end(); ++bi) {
// If exiting if or loop, update stacks
if ((*bi)->id() == currentMergeBlockId.top()) {
assume_branches_live.pop();
currentMergeBlockId.pop();
}
for (auto ii = (*bi)->begin(); ii != (*bi)->end(); ++ii) {
SpvOp op = ii->opcode();
switch (op) {
case SpvOpStore: {
uint32_t varId;
(void)GetPtr(&*ii, &varId);
// Mark stores as live if their variable is not function scope
// and is not private scope. Remember private stores for possible
// later inclusion. We cannot call IsLocalVar at this point because
// private_like_local_ has not been set yet.
if (IsVarOfStorage(varId, SpvStorageClassPrivate) ||
IsVarOfStorage(varId, SpvStorageClassWorkgroup))
private_stores_.push_back(&*ii);
else if (!IsVarOfStorage(varId, SpvStorageClassFunction))
AddToWorklist(&*ii);
} break;
case SpvOpCopyMemory:
case SpvOpCopyMemorySized: {
uint32_t varId;
(void)GetPtr(ii->GetSingleWordInOperand(kCopyMemoryTargetAddrInIdx),
&varId);
if (IsVarOfStorage(varId, SpvStorageClassPrivate) ||
IsVarOfStorage(varId, SpvStorageClassWorkgroup))
private_stores_.push_back(&*ii);
else if (!IsVarOfStorage(varId, SpvStorageClassFunction))
AddToWorklist(&*ii);
} break;
case SpvOpLoopMerge: {
assume_branches_live.push(false);
currentMergeBlockId.push(
ii->GetSingleWordInOperand(kLoopMergeMergeBlockIdInIdx));
} break;
case SpvOpSelectionMerge: {
assume_branches_live.push(false);
currentMergeBlockId.push(
ii->GetSingleWordInOperand(kSelectionMergeMergeBlockIdInIdx));
} break;
case SpvOpSwitch:
case SpvOpBranch:
case SpvOpBranchConditional:
case SpvOpUnreachable: {
if (assume_branches_live.top()) {
AddToWorklist(&*ii);
}
} break;
default: {
// Function calls, atomics, function params, function returns, etc.
// TODO(greg-lunarg): function calls live only if write to non-local
if (!ii->IsOpcodeSafeToDelete()) {
AddToWorklist(&*ii);
}
// Remember function calls
if (op == SpvOpFunctionCall) call_in_func_ = true;
} break;
}
}
}
// See if current function is an entry point
for (auto& ei : get_module()->entry_points()) {
if (ei.GetSingleWordInOperand(kEntryPointFunctionIdInIdx) ==
func->result_id()) {
func_is_entry_point_ = true;
break;
}
}
// If the current function is an entry point and has no function calls,
// we can optimize private variables as locals
private_like_local_ = func_is_entry_point_ && !call_in_func_;
// If privates are not like local, add their stores to worklist
if (!private_like_local_)
for (auto& ps : private_stores_) AddToWorklist(ps);
// Perform closure on live instruction set.
while (!worklist_.empty()) {
Instruction* liveInst = worklist_.front();
// Add all operand instructions if not already live
liveInst->ForEachInId([&liveInst, this](const uint32_t* iid) {
Instruction* inInst = get_def_use_mgr()->GetDef(*iid);
// Do not add label if an operand of a branch. This is not needed
// as part of live code discovery and can create false live code,
// for example, the branch to a header of a loop.
if (inInst->opcode() == SpvOpLabel && liveInst->IsBranch()) return;
AddToWorklist(inInst);
});
if (liveInst->type_id() != 0) {
AddToWorklist(get_def_use_mgr()->GetDef(liveInst->type_id()));
}
// If in a structured if or loop construct, add the controlling
// conditional branch and its merge.
BasicBlock* blk = context()->get_instr_block(liveInst);
Instruction* branchInst = block2headerBranch_[blk];
if (branchInst != nullptr) {
AddToWorklist(branchInst);
Instruction* mergeInst = branch2merge_[branchInst];
AddToWorklist(mergeInst);
}
// If the block is a header, add the next outermost controlling
// conditional branch and its merge.
Instruction* nextBranchInst = header2nextHeaderBranch_[blk];
if (nextBranchInst != nullptr) {
AddToWorklist(nextBranchInst);
Instruction* mergeInst = branch2merge_[nextBranchInst];
AddToWorklist(mergeInst);
}
// If local load, add all variable's stores if variable not already live
if (liveInst->opcode() == SpvOpLoad || liveInst->IsAtomicWithLoad()) {
uint32_t varId;
(void)GetPtr(liveInst, &varId);
if (varId != 0) {
ProcessLoad(varId);
}
// Process memory copies like loads
} else if (liveInst->opcode() == SpvOpCopyMemory ||
liveInst->opcode() == SpvOpCopyMemorySized) {
uint32_t varId;
(void)GetPtr(liveInst->GetSingleWordInOperand(kCopyMemorySourceAddrInIdx),
&varId);
if (varId != 0) {
ProcessLoad(varId);
}
// If merge, add other branches that are part of its control structure
} else if (liveInst->opcode() == SpvOpLoopMerge ||
liveInst->opcode() == SpvOpSelectionMerge) {
AddBreaksAndContinuesToWorklist(liveInst);
// If function call, treat as if it loads from all pointer arguments
} else if (liveInst->opcode() == SpvOpFunctionCall) {
liveInst->ForEachInId([this](const uint32_t* iid) {
// Skip non-ptr args
if (!IsPtr(*iid)) return;
uint32_t varId;
(void)GetPtr(*iid, &varId);
ProcessLoad(varId);
});
// If function parameter, treat as if it's result id is loaded from
} else if (liveInst->opcode() == SpvOpFunctionParameter) {
ProcessLoad(liveInst->result_id());
// We treat an OpImageTexelPointer as a load of the pointer, and
// that value is manipulated to get the result.
} else if (liveInst->opcode() == SpvOpImageTexelPointer) {
uint32_t varId;
(void)GetPtr(liveInst, &varId);
if (varId != 0) {
ProcessLoad(varId);
}
}
// Add OpDecorateId instructions that apply to this instruction to the work
// list. We use the decoration manager to look through the group
// decorations to get to the OpDecorate* instructions themselves.
auto decorations =
get_decoration_mgr()->GetDecorationsFor(liveInst->result_id(), false);
for (Instruction* dec : decorations) {
// We only care about OpDecorateId instructions because the are the only
// decorations that will reference an id that will have to be kept live
// because of that use.
if (dec->opcode() != SpvOpDecorateId) {
continue;
}
if (dec->GetSingleWordInOperand(1) ==
SpvDecorationHlslCounterBufferGOOGLE) {
// These decorations should not force the use id to be live. It will be
// removed if either the target or the in operand are dead.
continue;
}
AddToWorklist(dec);
}
worklist_.pop();
}
// Kill dead instructions and remember dead blocks
for (auto bi = structuredOrder.begin(); bi != structuredOrder.end();) {
uint32_t mergeBlockId = 0;
(*bi)->ForEachInst([this, &modified, &mergeBlockId](Instruction* inst) {
if (!IsDead(inst)) return;
if (inst->opcode() == SpvOpLabel) return;
// If dead instruction is selection merge, remember merge block
// for new branch at end of block
if (inst->opcode() == SpvOpSelectionMerge ||
inst->opcode() == SpvOpLoopMerge)
mergeBlockId = inst->GetSingleWordInOperand(0);
to_kill_.push_back(inst);
modified = true;
});
// If a structured if or loop was deleted, add a branch to its merge
// block, and traverse to the merge block and continue processing there.
// We know the block still exists because the label is not deleted.
if (mergeBlockId != 0) {
AddBranch(mergeBlockId, *bi);
for (++bi; (*bi)->id() != mergeBlockId; ++bi) {
}
auto merge_terminator = (*bi)->terminator();
if (merge_terminator->opcode() == SpvOpUnreachable) {
// The merge was unreachable. This is undefined behaviour so just
// return (or return an undef). Then mark the new return as live.
auto func_ret_type_inst = get_def_use_mgr()->GetDef(func->type_id());
if (func_ret_type_inst->opcode() == SpvOpTypeVoid) {
merge_terminator->SetOpcode(SpvOpReturn);
} else {
// Find an undef for the return value and make sure it gets kept by
// the pass.
auto undef_id = Type2Undef(func->type_id());
auto undef = get_def_use_mgr()->GetDef(undef_id);
live_insts_.Set(undef->unique_id());
merge_terminator->SetOpcode(SpvOpReturnValue);
merge_terminator->SetInOperands({{SPV_OPERAND_TYPE_ID, {undef_id}}});
get_def_use_mgr()->AnalyzeInstUse(merge_terminator);
}
live_insts_.Set(merge_terminator->unique_id());
}
} else {
++bi;
}
}
return modified;
}
void AggressiveDCEPass::InitializeModuleScopeLiveInstructions() {
// Keep all execution modes.
for (auto& exec : get_module()->execution_modes()) {
AddToWorklist(&exec);
}
// Keep all entry points.
for (auto& entry : get_module()->entry_points()) {
if (get_module()->version() >= SPV_SPIRV_VERSION_WORD(1, 4)) {
// In SPIR-V 1.4 and later, entry points must list all global variables
// used. DCE can still remove non-input/output variables and update the
// interface list. Mark the entry point as live and inputs and outputs as
// live, but defer decisions all other interfaces.
live_insts_.Set(entry.unique_id());
// The actual function is live always.
AddToWorklist(
get_def_use_mgr()->GetDef(entry.GetSingleWordInOperand(1u)));
for (uint32_t i = 3; i < entry.NumInOperands(); ++i) {
auto* var = get_def_use_mgr()->GetDef(entry.GetSingleWordInOperand(i));
auto storage_class = var->GetSingleWordInOperand(0u);
if (storage_class == SpvStorageClassInput ||
storage_class == SpvStorageClassOutput) {
AddToWorklist(var);
}
}
} else {
AddToWorklist(&entry);
}
}
for (auto& anno : get_module()->annotations()) {
if (anno.opcode() == SpvOpDecorate) {
// Keep workgroup size.
if (anno.GetSingleWordInOperand(1u) == SpvDecorationBuiltIn &&
anno.GetSingleWordInOperand(2u) == SpvBuiltInWorkgroupSize) {
AddToWorklist(&anno);
}
if (context()->preserve_bindings()) {
// Keep all bindings.
if ((anno.GetSingleWordInOperand(1u) == SpvDecorationDescriptorSet) ||
(anno.GetSingleWordInOperand(1u) == SpvDecorationBinding)) {
AddToWorklist(&anno);
}
}
if (context()->preserve_spec_constants()) {
// Keep all specialization constant instructions
if (anno.GetSingleWordInOperand(1u) == SpvDecorationSpecId) {
AddToWorklist(&anno);
}
}
}
}
}
Pass::Status AggressiveDCEPass::ProcessImpl() {
// Current functionality assumes shader capability
// TODO(greg-lunarg): Handle additional capabilities
if (!context()->get_feature_mgr()->HasCapability(SpvCapabilityShader))
return Status::SuccessWithoutChange;
// Current functionality assumes relaxed logical addressing (see
// instruction.h)
// TODO(greg-lunarg): Handle non-logical addressing
if (context()->get_feature_mgr()->HasCapability(SpvCapabilityAddresses))
return Status::SuccessWithoutChange;
// The variable pointer extension is no longer needed to use the capability,
// so we have to look for the capability.
if (context()->get_feature_mgr()->HasCapability(
SpvCapabilityVariablePointersStorageBuffer))
return Status::SuccessWithoutChange;
// If any extensions in the module are not explicitly supported,
// return unmodified.
if (!AllExtensionsSupported()) return Status::SuccessWithoutChange;
// Eliminate Dead functions.
bool modified = EliminateDeadFunctions();
InitializeModuleScopeLiveInstructions();
// Process all entry point functions.
ProcessFunction pfn = [this](Function* fp) { return AggressiveDCE(fp); };
modified |= context()->ProcessEntryPointCallTree(pfn);
// If the decoration manager is kept live then the context will try to keep it
// up to date. ADCE deals with group decorations by changing the operands in
// |OpGroupDecorate| instruction directly without informing the decoration
// manager. This can put it in an invalid state which will cause an error
// when the context tries to update it. To avoid this problem invalidate
// the decoration manager upfront.
//
// We kill it at now because it is used when processing the entry point
// functions.
context()->InvalidateAnalyses(IRContext::Analysis::kAnalysisDecorations);
// Process module-level instructions. Now that all live instructions have
// been marked, it is safe to remove dead global values.
modified |= ProcessGlobalValues();
// Sanity check.
assert(to_kill_.size() == 0 || modified);
// Kill all dead instructions.
for (auto inst : to_kill_) {
context()->KillInst(inst);
}
// Cleanup all CFG including all unreachable blocks.
ProcessFunction cleanup = [this](Function* f) { return CFGCleanup(f); };
modified |= context()->ProcessEntryPointCallTree(cleanup);
return modified ? Status::SuccessWithChange : Status::SuccessWithoutChange;
}
bool AggressiveDCEPass::EliminateDeadFunctions() {
// Identify live functions first. Those that are not live
// are dead. ADCE is disabled for non-shaders so we do not check for exported
// functions here.
std::unordered_set<const Function*> live_function_set;
ProcessFunction mark_live = [&live_function_set](Function* fp) {
live_function_set.insert(fp);
return false;
};
context()->ProcessEntryPointCallTree(mark_live);
bool modified = false;
for (auto funcIter = get_module()->begin();
funcIter != get_module()->end();) {
if (live_function_set.count(&*funcIter) == 0) {
modified = true;
EliminateFunction(&*funcIter);
funcIter = funcIter.Erase();
} else {
++funcIter;
}
}
return modified;
}
void AggressiveDCEPass::EliminateFunction(Function* func) {
// Remove all of the instruction in the function body
func->ForEachInst([this](Instruction* inst) { context()->KillInst(inst); },
true);
}
bool AggressiveDCEPass::ProcessGlobalValues() {
// Remove debug and annotation statements referencing dead instructions.
// This must be done before killing the instructions, otherwise there are
// dead objects in the def/use database.
bool modified = false;
Instruction* instruction = &*get_module()->debug2_begin();
while (instruction) {
if (instruction->opcode() != SpvOpName) {
instruction = instruction->NextNode();
continue;
}
if (IsTargetDead(instruction)) {
instruction = context()->KillInst(instruction);
modified = true;
} else {
instruction = instruction->NextNode();
}
}
// This code removes all unnecessary decorations safely (see #1174). It also
// does so in a more efficient manner than deleting them only as the targets
// are deleted.
std::vector<Instruction*> annotations;
for (auto& inst : get_module()->annotations()) annotations.push_back(&inst);
std::sort(annotations.begin(), annotations.end(), DecorationLess());
for (auto annotation : annotations) {
switch (annotation->opcode()) {
case SpvOpDecorate:
case SpvOpMemberDecorate:
case SpvOpDecorateStringGOOGLE:
case SpvOpMemberDecorateStringGOOGLE:
if (IsTargetDead(annotation)) {
context()->KillInst(annotation);
modified = true;
}
break;
case SpvOpDecorateId:
if (IsTargetDead(annotation)) {
context()->KillInst(annotation);
modified = true;
} else {
if (annotation->GetSingleWordInOperand(1) ==
SpvDecorationHlslCounterBufferGOOGLE) {
// HlslCounterBuffer will reference an id other than the target.
// If that id is dead, then the decoration can be removed as well.
uint32_t counter_buffer_id = annotation->GetSingleWordInOperand(2);
Instruction* counter_buffer_inst =
get_def_use_mgr()->GetDef(counter_buffer_id);
if (IsDead(counter_buffer_inst)) {
context()->KillInst(annotation);
modified = true;
}
}
}
break;
case SpvOpGroupDecorate: {
// Go through the targets of this group decorate. Remove each dead
// target. If all targets are dead, remove this decoration.
bool dead = true;
bool removed_operand = false;
for (uint32_t i = 1; i < annotation->NumOperands();) {
Instruction* opInst =
get_def_use_mgr()->GetDef(annotation->GetSingleWordOperand(i));
if (IsDead(opInst)) {
// Don't increment |i|.
annotation->RemoveOperand(i);
modified = true;
removed_operand = true;
} else {
i++;
dead = false;
}
}
if (dead) {
context()->KillInst(annotation);
modified = true;
} else if (removed_operand) {
context()->UpdateDefUse(annotation);
}
break;
}
case SpvOpGroupMemberDecorate: {
// Go through the targets of this group member decorate. Remove each
// dead target (and member index). If all targets are dead, remove this
// decoration.
bool dead = true;
bool removed_operand = false;
for (uint32_t i = 1; i < annotation->NumOperands();) {
Instruction* opInst =
get_def_use_mgr()->GetDef(annotation->GetSingleWordOperand(i));
if (IsDead(opInst)) {
// Don't increment |i|.
annotation->RemoveOperand(i + 1);
annotation->RemoveOperand(i);
modified = true;
removed_operand = true;
} else {
i += 2;
dead = false;
}
}
if (dead) {
context()->KillInst(annotation);
modified = true;
} else if (removed_operand) {
context()->UpdateDefUse(annotation);
}
break;
}
case SpvOpDecorationGroup:
// By the time we hit decoration groups we've checked everything that
// can target them. So if they have no uses they must be dead.
if (get_def_use_mgr()->NumUsers(annotation) == 0) {
context()->KillInst(annotation);
modified = true;
}
break;
default:
assert(false);
break;
}
}
// Since ADCE is disabled for non-shaders, we don't check for export linkage
// attributes here.
for (auto& val : get_module()->types_values()) {
if (IsDead(&val)) {
// Save forwarded pointer if pointer is live since closure does not mark
// this live as it does not have a result id. This is a little too
// conservative since it is not known if the structure type that needed
// it is still live. TODO(greg-lunarg): Only save if needed.
if (val.opcode() == SpvOpTypeForwardPointer) {
uint32_t ptr_ty_id = val.GetSingleWordInOperand(0);
Instruction* ptr_ty_inst = get_def_use_mgr()->GetDef(ptr_ty_id);
if (!IsDead(ptr_ty_inst)) continue;
}
to_kill_.push_back(&val);
modified = true;
}
}
if (get_module()->version() >= SPV_SPIRV_VERSION_WORD(1, 4)) {
// Remove the dead interface variables from the entry point interface list.
for (auto& entry : get_module()->entry_points()) {
std::vector<Operand> new_operands;
for (uint32_t i = 0; i < entry.NumInOperands(); ++i) {
if (i < 3) {
// Execution model, function id and name are always valid.
new_operands.push_back(entry.GetInOperand(i));
} else {
auto* var =
get_def_use_mgr()->GetDef(entry.GetSingleWordInOperand(i));
if (!IsDead(var)) {
new_operands.push_back(entry.GetInOperand(i));
}
}
}
if (new_operands.size() != entry.NumInOperands()) {
entry.SetInOperands(std::move(new_operands));
get_def_use_mgr()->UpdateDefUse(&entry);
}
}
}
return modified;
}
AggressiveDCEPass::AggressiveDCEPass() = default;
Pass::Status AggressiveDCEPass::Process() {
// Initialize extensions whitelist
InitExtensions();
return ProcessImpl();
}
void AggressiveDCEPass::InitExtensions() {
extensions_whitelist_.clear();
extensions_whitelist_.insert({
"SPV_AMD_shader_explicit_vertex_parameter",
"SPV_AMD_shader_trinary_minmax",
"SPV_AMD_gcn_shader",
"SPV_KHR_shader_ballot",
"SPV_AMD_shader_ballot",
"SPV_AMD_gpu_shader_half_float",
"SPV_KHR_shader_draw_parameters",
"SPV_KHR_subgroup_vote",
"SPV_KHR_16bit_storage",
"SPV_KHR_device_group",
"SPV_KHR_multiview",
"SPV_NVX_multiview_per_view_attributes",
"SPV_NV_viewport_array2",
"SPV_NV_stereo_view_rendering",
"SPV_NV_sample_mask_override_coverage",
"SPV_NV_geometry_shader_passthrough",
"SPV_AMD_texture_gather_bias_lod",
"SPV_KHR_storage_buffer_storage_class",
// SPV_KHR_variable_pointers
// Currently do not support extended pointer expressions
"SPV_AMD_gpu_shader_int16",
"SPV_KHR_post_depth_coverage",
"SPV_KHR_shader_atomic_counter_ops",
2018-03-08 13:54:00 +00:00
"SPV_EXT_shader_stencil_export",
"SPV_EXT_shader_viewport_index_layer",
"SPV_AMD_shader_image_load_store_lod",
"SPV_AMD_shader_fragment_mask",
"SPV_EXT_fragment_fully_covered",
"SPV_AMD_gpu_shader_half_float_fetch",
"SPV_GOOGLE_decorate_string",
"SPV_GOOGLE_hlsl_functionality1",
"SPV_GOOGLE_user_type",
"SPV_NV_shader_subgroup_partitioned",
"SPV_EXT_descriptor_indexing",
"SPV_NV_fragment_shader_barycentric",
"SPV_NV_compute_shader_derivatives",
"SPV_NV_shader_image_footprint",
"SPV_NV_shading_rate",
"SPV_NV_mesh_shader",
"SPV_NV_ray_tracing",
"SPV_EXT_fragment_invocation_density",
"SPV_EXT_physical_storage_buffer",
});
}
} // namespace opt
} // namespace spvtools