AuroraMiddleware/SPIRV-Tools
1
0
Fork 0
mirror of https://github.com/KhronosGroup/SPIRV-Tools synced 2024-11-27 05:40:06 +00:00
Code Issues Packages Projects Releases Wiki Activity

Revert "Revert "[spirv-opt] refactor inlining pass (#3328)" (#3342)" (#3345)

Browse Source 
This reverts commit d4fac3451b.
This commit is contained in:
Steven Perron 2020-05-14 10:55:47 -04:00 committed by GitHub
parent 3118276370
commit bd0a2da946
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
4 changed files with 1075 additions and 1086 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

603
source/opt/inline_pass.cpp
View File

@ -20,6 +20,7 @@
#include <utility>
#include "source/cfa.h"
#include "source/opt/reflect.h"
#include "source/util/make_unique.h"
// Indices of operands in SPIR-V instructions
@ -232,161 +233,10 @@ bool InlinePass::CloneSameBlockOps(
});
}
bool InlinePass::GenInlineCode(
std::vector<std::unique_ptr<BasicBlock>>* new_blocks,
std::vector<std::unique_ptr<Instruction>>* new_vars,
BasicBlock::iterator call_inst_itr,
void InlinePass::MoveInstsBeforeEntryBlock(
std::unordered_map<uint32_t, Instruction*>* preCallSB,
BasicBlock* new_blk_ptr, BasicBlock::iterator call_inst_itr,
UptrVectorIterator<BasicBlock> call_block_itr) {
// Map from all ids in the callee to their equivalent id in the caller
// as callee instructions are copied into caller.
std::unordered_map<uint32_t, uint32_t> callee2caller;
// Pre-call same-block insts
std::unordered_map<uint32_t, Instruction*> preCallSB;
// Post-call same-block op ids
std::unordered_map<uint32_t, uint32_t> postCallSB;
// Invalidate the def-use chains. They are not kept up to date while
// inlining. However, certain calls try to keep them up-to-date if they are
// valid. These operations can fail.
context()->InvalidateAnalyses(IRContext::kAnalysisDefUse);
Function* calleeFn = id2function_[call_inst_itr->GetSingleWordOperand(
kSpvFunctionCallFunctionId)];
// Check for multiple returns in the callee.
auto fi = early_return_funcs_.find(calleeFn->result_id());
const bool earlyReturn = fi != early_return_funcs_.end();
// Map parameters to actual arguments.
MapParams(calleeFn, call_inst_itr, &callee2caller);
// Define caller local variables for all callee variables and create map to
// them.
if (!CloneAndMapLocals(calleeFn, new_vars, &callee2caller)) {
return false;
}
// Create return var if needed.
const uint32_t calleeTypeId = calleeFn->type_id();
uint32_t returnVarId = 0;
analysis::Type* calleeType = context()->get_type_mgr()->GetType(calleeTypeId);
if (calleeType->AsVoid() == nullptr) {
returnVarId = CreateReturnVar(calleeFn, new_vars);
if (returnVarId == 0) {
return false;
}
}
// Create set of callee result ids. Used to detect forward references
std::unordered_set<uint32_t> callee_result_ids;
calleeFn->ForEachInst([&callee_result_ids](const Instruction* cpi) {
const uint32_t rid = cpi->result_id();
if (rid != 0) callee_result_ids.insert(rid);
});
// If the caller is a loop header and the callee has multiple blocks, then the
// normal inlining logic will place the OpLoopMerge in the last of several
// blocks in the loop. Instead, it should be placed at the end of the first
// block. We'll wait to move the OpLoopMerge until the end of the regular
// inlining logic, and only if necessary.
bool caller_is_loop_header = false;
if (call_block_itr->GetLoopMergeInst()) {
caller_is_loop_header = true;
}
bool callee_begins_with_structured_header =
(*(calleeFn->begin())).GetMergeInst() != nullptr;
// Clone and map callee code. Copy caller block code to beginning of
// first block and end of last block.
bool prevInstWasReturn = false;
uint32_t singleTripLoopHeaderId = 0;
uint32_t singleTripLoopContinueId = 0;
uint32_t returnLabelId = 0;
bool multiBlocks = false;
// new_blk_ptr is a new basic block in the caller. New instructions are
// written to it. It is created when we encounter the OpLabel
// of the first callee block. It is appended to new_blocks only when
// it is complete.
std::unique_ptr<BasicBlock> new_blk_ptr;
bool successful = calleeFn->WhileEachInst(
[&new_blocks, &callee2caller, &call_block_itr, &call_inst_itr,
&new_blk_ptr, &prevInstWasReturn, &returnLabelId, &returnVarId,
caller_is_loop_header, callee_begins_with_structured_header,
&calleeTypeId, &multiBlocks, &postCallSB, &preCallSB, earlyReturn,
&singleTripLoopHeaderId, &singleTripLoopContinueId, &callee_result_ids,
this](const Instruction* cpi) {
switch (cpi->opcode()) {
case SpvOpFunction:
case SpvOpFunctionParameter:
// Already processed
break;
case SpvOpVariable:
if (cpi->NumInOperands() == 2) {
assert(callee2caller.count(cpi->result_id()) &&
"Expected the variable to have already been mapped.");
uint32_t new_var_id = callee2caller.at(cpi->result_id());
// The initializer must be a constant or global value. No mapped
// should be used.
uint32_t val_id = cpi->GetSingleWordInOperand(1);
AddStore(new_var_id, val_id, &new_blk_ptr);
}
break;
case SpvOpUnreachable:
case SpvOpKill: {
// Generate a return label so that we split the block with the
// function call. Copy the terminator into the new block.
if (returnLabelId == 0) {
returnLabelId = context()->TakeNextId();
if (returnLabelId == 0) {
return false;
}
}
std::unique_ptr<Instruction> terminator(
new Instruction(context(), cpi->opcode(), 0, 0, {}));
new_blk_ptr->AddInstruction(std::move(terminator));
break;
}
case SpvOpLabel: {
// If previous instruction was early return, insert branch
// instruction to return block.
if (prevInstWasReturn) {
if (returnLabelId == 0) {
returnLabelId = context()->TakeNextId();
if (returnLabelId == 0) {
return false;
}
}
AddBranch(returnLabelId, &new_blk_ptr);
prevInstWasReturn = false;
}
// Finish current block (if it exists) and get label for next block.
uint32_t labelId;
bool firstBlock = false;
if (new_blk_ptr != nullptr) {
new_blocks->push_back(std::move(new_blk_ptr));
// If result id is already mapped, use it, otherwise get a new
// one.
const uint32_t rid = cpi->result_id();
const auto mapItr = callee2caller.find(rid);
labelId = (mapItr != callee2caller.end())
? mapItr->second
: context()->TakeNextId();
if (labelId == 0) {
return false;
}
} else {
// First block needs to use label of original block
// but map callee label in case of phi reference.
labelId = call_block_itr->id();
callee2caller[cpi->result_id()] = labelId;
firstBlock = true;
}
// Create first/next block.
new_blk_ptr = MakeUnique<BasicBlock>(NewLabel(labelId));
if (firstBlock) {
// Copy contents of original caller block up to call instruction.
for (auto cii = call_block_itr->begin(); cii != call_inst_itr;
cii = call_block_itr->begin()) {
Instruction* inst = &*cii;
@ -395,22 +245,19 @@ bool InlinePass::GenInlineCode(
// Remember same-block ops for possible regeneration.
if (IsSameBlockOp(&*cp_inst)) {
auto* sb_inst_ptr = cp_inst.get();
preCallSB[cp_inst->result_id()] = sb_inst_ptr;
(*preCallSB)[cp_inst->result_id()] = sb_inst_ptr;
}
new_blk_ptr->AddInstruction(std::move(cp_inst));
}
if (caller_is_loop_header &&
callee_begins_with_structured_header) {
// We can't place both the caller's merge instruction and
// another merge instruction in the same block. So split the
// calling block. Insert an unconditional branch to a new guard
// block. Later, once we know the ID of the last block, we
// will move the caller's OpLoopMerge from the last generated
// block into the first block. We also wait to avoid
// invalidating various iterators.
}
std::unique_ptr<BasicBlock> InlinePass::AddGuardBlock(
std::vector<std::unique_ptr<BasicBlock>>* new_blocks,
std::unordered_map<uint32_t, uint32_t>* callee2caller,
std::unique_ptr<BasicBlock> new_blk_ptr, uint32_t entry_blk_label_id) {
const auto guard_block_id = context()->TakeNextId();
if (guard_block_id == 0) {
return false;
return nullptr;
}
AddBranch(guard_block_id, &new_blk_ptr);
new_blocks->push_back(std::move(new_blk_ptr));
@ -419,104 +266,143 @@ bool InlinePass::GenInlineCode(
// Reset the mapping of the callee's entry block to point to
// the guard block. Do this so we can fix up phis later on to
// satisfy dominance.
callee2caller[cpi->result_id()] = guard_block_id;
(*callee2caller)[entry_blk_label_id] = guard_block_id;
return new_blk_ptr;
}
// If callee has early return, insert a header block for
// single-trip loop that will encompass callee code. Start
// postheader block.
//
// Note: Consider the following combination:
// - the caller is a single block loop
// - the callee does not begin with a structure header
// - the callee has multiple returns.
// We still need to split the caller block and insert a guard
// block. But we only need to do it once. We haven't done it yet,
// but the single-trip loop header will serve the same purpose.
if (earlyReturn) {
singleTripLoopHeaderId = context()->TakeNextId();
if (singleTripLoopHeaderId == 0) {
return false;
InstructionList::iterator InlinePass::AddStoresForVariableInitializers(
const std::unordered_map<uint32_t, uint32_t>& callee2caller,
std::unique_ptr<BasicBlock>* new_blk_ptr,
UptrVectorIterator<BasicBlock> callee_first_block_itr) {
auto callee_var_itr = callee_first_block_itr->begin();
while (callee_var_itr->opcode() == SpvOp::SpvOpVariable) {
if (callee_var_itr->NumInOperands() == 2) {
assert(callee2caller.count(callee_var_itr->result_id()) &&
"Expected the variable to have already been mapped.");
uint32_t new_var_id = callee2caller.at(callee_var_itr->result_id());
// The initializer must be a constant or global value. No mapped
// should be used.
uint32_t val_id = callee_var_itr->GetSingleWordInOperand(1);
AddStore(new_var_id, val_id, new_blk_ptr);
}
AddBranch(singleTripLoopHeaderId, &new_blk_ptr);
new_blocks->push_back(std::move(new_blk_ptr));
new_blk_ptr =
MakeUnique<BasicBlock>(NewLabel(singleTripLoopHeaderId));
returnLabelId = context()->TakeNextId();
singleTripLoopContinueId = context()->TakeNextId();
if (returnLabelId == 0 || singleTripLoopContinueId == 0) {
return false;
++callee_var_itr;
}
AddLoopMerge(returnLabelId, singleTripLoopContinueId,
&new_blk_ptr);
uint32_t postHeaderId = context()->TakeNextId();
if (postHeaderId == 0) {
return false;
return callee_var_itr;
}
AddBranch(postHeaderId, &new_blk_ptr);
new_blocks->push_back(std::move(new_blk_ptr));
new_blk_ptr = MakeUnique<BasicBlock>(NewLabel(postHeaderId));
multiBlocks = true;
// Reset the mapping of the callee's entry block to point to
// the post-header block. Do this so we can fix up phis later
// on to satisfy dominance.
callee2caller[cpi->result_id()] = postHeaderId;
bool InlinePass::InlineInstructionInBB(
const std::unordered_map<uint32_t, uint32_t>& callee2caller,
BasicBlock* new_blk_ptr, const Instruction* inst) {
// If we have return, it must be at the end of the callee. We will handle
// it at the end.
if (inst->opcode() == SpvOpReturnValue || inst->opcode() == SpvOpReturn)
return true;
// Copy callee instruction and remap all input Ids.
std::unique_ptr<Instruction> cp_inst(inst->Clone(context()));
cp_inst->ForEachInId([&callee2caller](uint32_t* iid) {
const auto mapItr = callee2caller.find(*iid);
if (mapItr != callee2caller.end()) {
*iid = mapItr->second;
}
} else {
multiBlocks = true;
});
// If result id is non-zero, remap it.
const uint32_t rid = cp_inst->result_id();
if (rid != 0) {
const auto mapItr = callee2caller.find(rid);
if (mapItr == callee2caller.end()) return false;
uint32_t nid = mapItr->second;
cp_inst->SetResultId(nid);
get_decoration_mgr()->CloneDecorations(rid, nid);
}
} break;
case SpvOpReturnValue: {
new_blk_ptr->AddInstruction(std::move(cp_inst));
return true;
}
std::unique_ptr<BasicBlock> InlinePass::InlineReturn(
const std::unordered_map<uint32_t, uint32_t>& callee2caller,
std::vector<std::unique_ptr<BasicBlock>>* new_blocks,
std::unique_ptr<BasicBlock> new_blk_ptr, Function* calleeFn,
const Instruction* inst, uint32_t returnVarId) {
// Store return value to return variable.
if (inst->opcode() == SpvOpReturnValue) {
assert(returnVarId != 0);
uint32_t valId = cpi->GetInOperand(kSpvReturnValueId).words[0];
uint32_t valId = inst->GetInOperand(kSpvReturnValueId).words[0];
const auto mapItr = callee2caller.find(valId);
if (mapItr != callee2caller.end()) {
valId = mapItr->second;
}
AddStore(returnVarId, valId, &new_blk_ptr);
}
// Remember we saw a return; if followed by a label, will need to
// insert branch.
prevInstWasReturn = true;
} break;
case SpvOpReturn: {
// Remember we saw a return; if followed by a label, will need to
// insert branch.
prevInstWasReturn = true;
} break;
case SpvOpFunctionEnd: {
// If there was an early return, we generated a return label id
// for it. Now we have to generate the return block with that Id.
if (returnLabelId != 0) {
// If previous instruction was return, insert branch instruction
// to return block.
if (prevInstWasReturn) AddBranch(returnLabelId, &new_blk_ptr);
if (earlyReturn) {
// If we generated a loop header for the single-trip loop
// to accommodate early returns, insert the continue
// target block now, with a false branch back to the loop
// header.
uint32_t returnLabelId = 0;
for (auto callee_block_itr = calleeFn->begin();
callee_block_itr != calleeFn->end(); ++callee_block_itr) {
if (callee_block_itr->tail()->opcode() == SpvOpUnreachable ||
callee_block_itr->tail()->opcode() == SpvOpKill) {
returnLabelId = context()->TakeNextId();
break;
}
}
if (returnLabelId == 0) return new_blk_ptr;
if (inst->opcode() == SpvOpReturn || inst->opcode() == SpvOpReturnValue)
AddBranch(returnLabelId, &new_blk_ptr);
new_blocks->push_back(std::move(new_blk_ptr));
new_blk_ptr =
MakeUnique<BasicBlock>(NewLabel(singleTripLoopContinueId));
uint32_t false_id = GetFalseId();
if (false_id == 0) {
return MakeUnique<BasicBlock>(NewLabel(returnLabelId));
}
bool InlinePass::InlineEntryBlock(
const std::unordered_map<uint32_t, uint32_t>& callee2caller,
std::unique_ptr<BasicBlock>* new_blk_ptr,
UptrVectorIterator<BasicBlock> callee_first_block) {
auto callee_inst_itr = AddStoresForVariableInitializers(
callee2caller, new_blk_ptr, callee_first_block);
while (callee_inst_itr != callee_first_block->end()) {
if (!InlineInstructionInBB(callee2caller, new_blk_ptr->get(),
&*callee_inst_itr)) {
return false;
}
AddBranchCond(false_id, singleTripLoopHeaderId, returnLabelId,
&new_blk_ptr);
++callee_inst_itr;
}
// Generate the return block.
return true;
}
std::unique_ptr<BasicBlock> InlinePass::InlineBasicBlocks(
std::vector<std::unique_ptr<BasicBlock>>* new_blocks,
const std::unordered_map<uint32_t, uint32_t>& callee2caller,
std::unique_ptr<BasicBlock> new_blk_ptr, Function* calleeFn) {
auto callee_block_itr = calleeFn->begin();
++callee_block_itr;
while (callee_block_itr != calleeFn->end()) {
new_blocks->push_back(std::move(new_blk_ptr));
new_blk_ptr = MakeUnique<BasicBlock>(NewLabel(returnLabelId));
multiBlocks = true;
const auto mapItr =
callee2caller.find(callee_block_itr->GetLabelInst()->result_id());
if (mapItr == callee2caller.end()) return nullptr;
new_blk_ptr = MakeUnique<BasicBlock>(NewLabel(mapItr->second));
auto tail_inst_itr = callee_block_itr->end();
for (auto inst_itr = callee_block_itr->begin(); inst_itr != tail_inst_itr;
++inst_itr) {
if (!InlineInstructionInBB(callee2caller, new_blk_ptr.get(),
&*inst_itr)) {
return nullptr;
}
// Load return value into result id of call, if it exists.
if (returnVarId != 0) {
const uint32_t resId = call_inst_itr->result_id();
assert(resId != 0);
AddLoad(calleeTypeId, resId, returnVarId, &new_blk_ptr);
}
++callee_block_itr;
}
return new_blk_ptr;
}
bool InlinePass::MoveCallerInstsAfterFunctionCall(
std::unordered_map<uint32_t, Instruction*>* preCallSB,
std::unordered_map<uint32_t, uint32_t>* postCallSB,
std::unique_ptr<BasicBlock>* new_blk_ptr,
BasicBlock::iterator call_inst_itr, bool multiBlocks) {
// Copy remaining instructions from caller block.
for (Instruction* inst = call_inst_itr->NextNode(); inst;
inst = call_inst_itr->NextNode()) {
@ -525,75 +411,24 @@ bool InlinePass::GenInlineCode(
// If multiple blocks generated, regenerate any same-block
// instruction that has not been seen in this last block.
if (multiBlocks) {
if (!CloneSameBlockOps(&cp_inst, &postCallSB, &preCallSB,
&new_blk_ptr)) {
if (!CloneSameBlockOps(&cp_inst, postCallSB, preCallSB, new_blk_ptr)) {
return false;
}
// Remember same-block ops in this block.
if (IsSameBlockOp(&*cp_inst)) {
const uint32_t rid = cp_inst->result_id();
postCallSB[rid] = rid;
(*postCallSB)[rid] = rid;
}
}
new_blk_ptr->AddInstruction(std::move(cp_inst));
}
// Finalize inline code.
new_blocks->push_back(std::move(new_blk_ptr));
} break;
default: {
// Copy callee instruction and remap all input Ids.
std::unique_ptr<Instruction> cp_inst(cpi->Clone(context()));
bool succeeded = cp_inst->WhileEachInId(
[&callee2caller, &callee_result_ids, this](uint32_t* iid) {
const auto mapItr = callee2caller.find(*iid);
if (mapItr != callee2caller.end()) {
*iid = mapItr->second;
} else if (callee_result_ids.find(*iid) !=
callee_result_ids.end()) {
// Forward reference. Allocate a new id, map it,
// use it and check for it when remapping result ids
const uint32_t nid = context()->TakeNextId();
if (nid == 0) {
return false;
}
callee2caller[*iid] = nid;
*iid = nid;
}
return true;
});
if (!succeeded) {
return false;
}
// If result id is non-zero, remap it. If already mapped, use mapped
// value, else use next id.
const uint32_t rid = cp_inst->result_id();
if (rid != 0) {
const auto mapItr = callee2caller.find(rid);
uint32_t nid;
if (mapItr != callee2caller.end()) {
nid = mapItr->second;
} else {
nid = context()->TakeNextId();
if (nid == 0) {
return false;
}
callee2caller[rid] = nid;
}
cp_inst->SetResultId(nid);
get_decoration_mgr()->CloneDecorations(rid, nid);
}
new_blk_ptr->AddInstruction(std::move(cp_inst));
} break;
}
return true;
});
if (!successful) {
return false;
new_blk_ptr->get()->AddInstruction(std::move(cp_inst));
}
if (caller_is_loop_header && (new_blocks->size() > 1)) {
return true;
}
void InlinePass::MoveLoopMergeInstToFirstBlock(
std::vector<std::unique_ptr<BasicBlock>>* new_blocks) {
// Move the OpLoopMerge from the last block back to the first, where
// it belongs.
auto& first = new_blocks->front();
@ -612,6 +447,126 @@ bool InlinePass::GenInlineCode(
delete &*loop_merge_itr;
}
bool InlinePass::GenInlineCode(
std::vector<std::unique_ptr<BasicBlock>>* new_blocks,
std::vector<std::unique_ptr<Instruction>>* new_vars,
BasicBlock::iterator call_inst_itr,
UptrVectorIterator<BasicBlock> call_block_itr) {
// Map from all ids in the callee to their equivalent id in the caller
// as callee instructions are copied into caller.
std::unordered_map<uint32_t, uint32_t> callee2caller;
// Pre-call same-block insts
std::unordered_map<uint32_t, Instruction*> preCallSB;
// Post-call same-block op ids
std::unordered_map<uint32_t, uint32_t> postCallSB;
// Invalidate the def-use chains. They are not kept up to date while
// inlining. However, certain calls try to keep them up-to-date if they are
// valid. These operations can fail.
context()->InvalidateAnalyses(IRContext::kAnalysisDefUse);
// If the caller is a loop header and the callee has multiple blocks, then the
// normal inlining logic will place the OpLoopMerge in the last of several
// blocks in the loop. Instead, it should be placed at the end of the first
// block. We'll wait to move the OpLoopMerge until the end of the regular
// inlining logic, and only if necessary.
bool caller_is_loop_header = call_block_itr->GetLoopMergeInst() != nullptr;
// Single-trip loop continue block
std::unique_ptr<BasicBlock> single_trip_loop_cont_blk;
Function* calleeFn = id2function_[call_inst_itr->GetSingleWordOperand(
kSpvFunctionCallFunctionId)];
// Map parameters to actual arguments.
MapParams(calleeFn, call_inst_itr, &callee2caller);
// Define caller local variables for all callee variables and create map to
// them.
if (!CloneAndMapLocals(calleeFn, new_vars, &callee2caller)) {
return false;
}
// First block needs to use label of original block
// but map callee label in case of phi reference.
uint32_t entry_blk_label_id = calleeFn->begin()->GetLabelInst()->result_id();
callee2caller[entry_blk_label_id] = call_block_itr->id();
std::unique_ptr<BasicBlock> new_blk_ptr =
MakeUnique<BasicBlock>(NewLabel(call_block_itr->id()));
// Move instructions of original caller block up to call instruction.
MoveInstsBeforeEntryBlock(&preCallSB, new_blk_ptr.get(), call_inst_itr,
call_block_itr);
if (caller_is_loop_header &&
(*(calleeFn->begin())).GetMergeInst() != nullptr) {
// We can't place both the caller's merge instruction and
// another merge instruction in the same block. So split the
// calling block. Insert an unconditional branch to a new guard
// block. Later, once we know the ID of the last block, we
// will move the caller's OpLoopMerge from the last generated
// block into the first block. We also wait to avoid
// invalidating various iterators.
new_blk_ptr = AddGuardBlock(new_blocks, &callee2caller,
std::move(new_blk_ptr), entry_blk_label_id);
if (new_blk_ptr == nullptr) return false;
}
// Create return var if needed.
const uint32_t calleeTypeId = calleeFn->type_id();
uint32_t returnVarId = 0;
analysis::Type* calleeType = context()->get_type_mgr()->GetType(calleeTypeId);
if (calleeType->AsVoid() == nullptr) {
returnVarId = CreateReturnVar(calleeFn, new_vars);
if (returnVarId == 0) {
return false;
}
}
calleeFn->WhileEachInst([&callee2caller, this](const Instruction* cpi) {
// Create set of callee result ids. Used to detect forward references
const uint32_t rid = cpi->result_id();
if (rid != 0 && callee2caller.find(rid) == callee2caller.end()) {
const uint32_t nid = context()->TakeNextId();
if (nid == 0) return false;
callee2caller[rid] = nid;
}
return true;
});
// Inline the entry block of the callee function.
if (!InlineEntryBlock(callee2caller, &new_blk_ptr, calleeFn->begin())) {
return false;
}
// Inline blocks of the callee function other than the entry block.
new_blk_ptr = InlineBasicBlocks(new_blocks, callee2caller,
std::move(new_blk_ptr), calleeFn);
if (new_blk_ptr == nullptr) return false;
new_blk_ptr =
InlineReturn(callee2caller, new_blocks, std::move(new_blk_ptr), calleeFn,
&*(calleeFn->tail()->tail()), returnVarId);
// Load return value into result id of call, if it exists.
if (returnVarId != 0) {
const uint32_t resId = call_inst_itr->result_id();
assert(resId != 0);
AddLoad(calleeTypeId, resId, returnVarId, &new_blk_ptr);
}
// Move instructions of original caller block after call instruction.
if (!MoveCallerInstsAfterFunctionCall(&preCallSB, &postCallSB, &new_blk_ptr,
call_inst_itr,
calleeFn->begin() != calleeFn->end()))
return false;
// Finalize inline code.
new_blocks->push_back(std::move(new_blk_ptr));
if (caller_is_loop_header && (new_blocks->size() > 1))
MoveLoopMergeInstToFirstBlock(new_blocks);
// Update block map given replacement blocks.
for (auto& blk : *new_blocks) {
id2block_[blk->id()] = &*blk;
@ -624,7 +579,21 @@ bool InlinePass::IsInlinableFunctionCall(const Instruction* inst) {
const uint32_t calleeFnId =
inst->GetSingleWordOperand(kSpvFunctionCallFunctionId);
const auto ci = inlinable_.find(calleeFnId);
return ci != inlinable_.cend();
if (ci == inlinable_.cend()) return false;
if (early_return_funcs_.find(calleeFnId) != early_return_funcs_.end()) {
// We rely on the merge-return pass to handle the early return case
// in advance.
std::string message =
"The function '" + id2function_[calleeFnId]->DefInst().PrettyPrint() +
"' could not be inlined because the return instruction "
"is not at the end of the function. This could be fixed by "
"running merge-return before inlining.";
consumer()(SPV_MSG_WARNING, "", {0, 0, 0}, message.c_str());
return false;
}
return true;
}
void InlinePass::UpdateSucceedingPhis(
@ -645,26 +614,6 @@ void InlinePass::UpdateSucceedingPhis(
});
}
bool InlinePass::HasNoReturnInStructuredConstruct(Function* func) {
// If control not structured, do not do loop/return analysis
// TODO: Analyze returns in non-structured control flow
if (!context()->get_feature_mgr()->HasCapability(SpvCapabilityShader))
return false;
const auto structured_analysis = context()->GetStructuredCFGAnalysis();
// Search for returns in structured construct.
bool return_in_construct = false;
for (auto& blk : *func) {
auto terminal_ii = blk.cend();
--terminal_ii;
if (spvOpcodeIsReturn(terminal_ii->opcode()) &&
structured_analysis->ContainingConstruct(blk.id()) != 0) {
return_in_construct = true;
break;
}
}
return !return_in_construct;
}
bool InlinePass::HasNoReturnInLoop(Function* func) {
// If control not structured, do not do loop/return analysis
// TODO: Analyze returns in non-structured control flow
@ -686,10 +635,18 @@ bool InlinePass::HasNoReturnInLoop(Function* func) {
}
void InlinePass::AnalyzeReturns(Function* func) {
// Analyze functions without a return in loop.
if (HasNoReturnInLoop(func)) {
no_return_in_loop_.insert(func->result_id());
if (!HasNoReturnInStructuredConstruct(func))
}
// Analyze functions with a return before its tail basic block.
for (auto& blk : *func) {
auto terminal_ii = blk.cend();
--terminal_ii;
if (spvOpcodeIsReturn(terminal_ii->opcode()) && &blk != func->tail()) {
early_return_funcs_.insert(func->result_id());
break;
}
}
}

62
source/opt/inline_pass.h
View File

@ -124,10 +124,6 @@ class InlinePass : public Pass {
// Return true if |inst| is a function call that can be inlined.
bool IsInlinableFunctionCall(const Instruction* inst);
// Return true if |func| does not have a return that is
// nested in a structured if, switch or loop.
bool HasNoReturnInStructuredConstruct(Function* func);
// Return true if |func| has no return in a loop. The current analysis
// requires structured control flow, so return false if control flow not
// structured ie. module is not a shader.
@ -171,6 +167,64 @@ class InlinePass : public Pass {
// Set of functions that are originally called directly or indirectly from a
// continue construct.
std::unordered_set<uint32_t> funcs_called_from_continue_;
private:
// Moves instructions of the caller function up to the call instruction
// to |new_blk_ptr|.
void MoveInstsBeforeEntryBlock(
std::unordered_map<uint32_t, Instruction*>* preCallSB,
BasicBlock* new_blk_ptr, BasicBlock::iterator call_inst_itr,
UptrVectorIterator<BasicBlock> call_block_itr);
// Returns a new guard block after adding a branch to the end of
// |new_blocks|.
std::unique_ptr<BasicBlock> AddGuardBlock(
std::vector<std::unique_ptr<BasicBlock>>* new_blocks,
std::unordered_map<uint32_t, uint32_t>* callee2caller,
std::unique_ptr<BasicBlock> new_blk_ptr, uint32_t entry_blk_label_id);
// Add store instructions for initializers of variables.
InstructionList::iterator AddStoresForVariableInitializers(
const std::unordered_map<uint32_t, uint32_t>& callee2caller,
std::unique_ptr<BasicBlock>* new_blk_ptr,
UptrVectorIterator<BasicBlock> callee_block_itr);
// Inlines a single instruction of the callee function.
bool InlineInstructionInBB(
const std::unordered_map<uint32_t, uint32_t>& callee2caller,
BasicBlock* new_blk_ptr, const Instruction* inst);
// Inlines the return instruction of the callee function.
std::unique_ptr<BasicBlock> InlineReturn(
const std::unordered_map<uint32_t, uint32_t>& callee2caller,
std::vector<std::unique_ptr<BasicBlock>>* new_blocks,
std::unique_ptr<BasicBlock> new_blk_ptr, Function* calleeFn,
const Instruction* inst, uint32_t returnVarId);
// Inlines the entry block of the callee function.
bool InlineEntryBlock(
const std::unordered_map<uint32_t, uint32_t>& callee2caller,
std::unique_ptr<BasicBlock>* new_blk_ptr,
UptrVectorIterator<BasicBlock> callee_first_block);
// Inlines basic blocks of the callee function other than the entry basic
// block.
std::unique_ptr<BasicBlock> InlineBasicBlocks(
std::vector<std::unique_ptr<BasicBlock>>* new_blocks,
const std::unordered_map<uint32_t, uint32_t>& callee2caller,
std::unique_ptr<BasicBlock> new_blk_ptr, Function* calleeFn);
// Moves instructions of the caller function after the call instruction
// to |new_blk_ptr|.
bool MoveCallerInstsAfterFunctionCall(
std::unordered_map<uint32_t, Instruction*>* preCallSB,
std::unordered_map<uint32_t, uint32_t>* postCallSB,
std::unique_ptr<BasicBlock>* new_blk_ptr,
BasicBlock::iterator call_inst_itr, bool multiBlocks);
// Move the OpLoopMerge from the last block back to the first.
void MoveLoopMergeInstToFirstBlock(
std::vector<std::unique_ptr<BasicBlock>>* new_blocks);
};
} // namespace opt

32
test/opt/inline_opaque_test.cpp
View File

@ -102,12 +102,12 @@ OpStore %30 %29
OpStore %32 %31
%33 = OpLoad %S_t %s0
OpStore %param %33
%41 = OpAccessChain %_ptr_Function_18 %param %int_2
%42 = OpLoad %18 %41
%43 = OpAccessChain %_ptr_Function_v2float %param %int_0
%44 = OpLoad %v2float %43
%45 = OpImageSampleImplicitLod %v4float %42 %44
OpStore %outColor %45
%42 = OpAccessChain %_ptr_Function_18 %param %int_2
%43 = OpLoad %18 %42
%44 = OpAccessChain %_ptr_Function_v2float %param %int_0
%45 = OpLoad %v2float %44
%46 = OpImageSampleImplicitLod %v4float %43 %45
OpStore %outColor %46
OpReturn
OpFunctionEnd
)";
@ -191,10 +191,10 @@ OpFunctionEnd
%34 = OpVariable %_ptr_Function_20 Function
%35 = OpVariable %_ptr_Function_20 Function
%25 = OpVariable %_ptr_Function_20 Function
%36 = OpLoad %20 %sampler16
OpStore %34 %36
%37 = OpLoad %20 %34
OpStore %35 %37
%37 = OpLoad %20 %sampler16
OpStore %34 %37
%38 = OpLoad %20 %34
OpStore %35 %38
%26 = OpLoad %20 %35
OpStore %25 %26
%27 = OpLoad %20 %25
@ -301,12 +301,12 @@ OpStore %31 %30
OpStore %33 %32
%34 = OpLoad %S_t %s0
OpStore %param %34
%44 = OpAccessChain %_ptr_Function_19 %param %int_2
%45 = OpLoad %19 %44
%46 = OpAccessChain %_ptr_Function_v2float %param %int_0
%47 = OpLoad %v2float %46
%48 = OpImageSampleImplicitLod %v4float %45 %47
OpStore %outColor %48
%45 = OpAccessChain %_ptr_Function_19 %param %int_2
%46 = OpLoad %19 %45
%47 = OpAccessChain %_ptr_Function_v2float %param %int_0
%48 = OpLoad %v2float %47
%49 = OpImageSampleImplicitLod %v4float %46 %48
OpStore %outColor %49
OpReturn
OpFunctionEnd
)";

1408
test/opt/inline_test.cpp
View File

File diff suppressed because it is too large Load Diff