SSATmp* IRBuilder::optimizeInst(IRInstruction* inst, CloneFlag doClone) {
if (auto const tmp = optimizeWork(inst, folly::none)) {
return tmp;
}
// Couldn't CSE or simplify the instruction; clone it and append.
if (inst->op() != Nop) {
if (doClone == CloneFlag::Yes) inst = m_unit.cloneInstruction(inst);
appendInstruction(inst);
// returns nullptr if instruction has no dest, returns the first
// (possibly only) dest otherwise
return inst->dst(0);
}
return nullptr;
}
SSATmp* TraceBuilder::optimizeInst(IRInstruction* inst) {
if (SSATmp* tmp = optimizeWork(inst)) {
return tmp;
}
// Couldn't CSE or simplify the instruction; clone it and append.
if (inst->op() != Nop) {
inst = inst->clone(&m_irFactory);
appendInstruction(inst);
// returns nullptr if instruction has no dest, returns the first
// (possibly only) dest otherwise
return inst->getDst(0);
}
return nullptr;
}
/*
* reoptimize() runs a trace through a second pass of IRBuilder
* optimizations, like this:
*
* reset state.
* move all blocks to a temporary list.
* compute immediate dominators.
* for each block in trace order:
* if we have a snapshot state for this block:
* clear cse entries that don't dominate this block.
* use snapshot state.
* move all instructions to a temporary list.
* for each instruction:
* optimizeWork - do CSE and simplify again
* if not simplified:
* append existing instruction and update state.
* else:
* if the instruction has a result, insert a mov from the
* simplified tmp to the original tmp and discard the instruction.
* if the last conditional branch was turned into a jump, remove the
* fall-through edge to the next block.
*/
void IRBuilder::reoptimize() {
Timer _t("optimize_reoptimize");
FTRACE(5, "ReOptimize:vvvvvvvvvvvvvvvvvvvv\n");
SCOPE_EXIT { FTRACE(5, "ReOptimize:^^^^^^^^^^^^^^^^^^^^\n"); };
always_assert(m_savedBlocks.empty());
always_assert(!m_curWhere);
always_assert(m_state.inlineDepth() == 0);
m_state.setEnableCse(RuntimeOption::EvalHHIRCse);
m_enableSimplification = RuntimeOption::EvalHHIRSimplification;
if (!m_state.enableCse() && !m_enableSimplification) return;
setConstrainGuards(false);
auto blocksIds = rpoSortCfgWithIds(m_unit);
auto const idoms = findDominators(m_unit, blocksIds);
m_state.clear();
for (auto* block : blocksIds.blocks) {
FTRACE(5, "Block: {}\n", block->id());
m_state.startBlock(block);
m_curBlock = block;
auto nextBlock = block->next();
auto backMarker = block->back().marker();
auto instructions = block->moveInstrs();
assert(block->empty());
while (!instructions.empty()) {
auto* inst = &instructions.front();
instructions.pop_front();
// merging state looks at the current marker, and optimizeWork
// below may create new instructions. Use the marker from this
// instruction.
assert(inst->marker().valid());
setMarker(inst->marker());
auto const tmp = optimizeWork(inst, idoms); // Can generate new instrs!
if (!tmp) {
// Could not optimize; keep the old instruction
appendInstruction(inst);
continue;
}
SSATmp* dst = inst->dst();
if (dst != tmp) {
// The result of optimization has a different destination than the inst.
// Generate a mov(tmp->dst) to get result into dst. If we get here then
// assume the last instruction in the block isn't a guard. If it was,
// we would have to insert the mov on the fall-through edge.
assert(block->empty() || !block->back().isBlockEnd());
appendInstruction(m_unit.mov(dst, tmp, inst->marker()));
}
if (inst->isBlockEnd()) {
// We're not re-adding the block-end instruction. Unset its edges.
inst->setTaken(nullptr);
inst->setNext(nullptr);
}
}
if (block->empty() || !block->back().isBlockEnd()) {
// Our block-end instruction was eliminated (most likely a Jmp* converted
// to a nop). Replace it with a jump to the next block.
appendInstruction(m_unit.gen(Jmp, backMarker, nextBlock));
}
m_state.finishBlock(block);
}
}
/*
* reoptimize() runs a trace through a second pass of TraceBuilder
* optimizations, like this:
*
* reset state.
* move all blocks to a temporary list.
* compute immediate dominators.
* for each block in trace order:
* if we have a snapshot state for this block:
* clear cse entries that don't dominate this block.
* use snapshot state.
* move all instructions to a temporary list.
* for each instruction:
* optimizeWork - do CSE and simplify again
* if not simplified:
* append existing instruction and update state.
* else:
* if the instruction has a result, insert a mov from the
* simplified tmp to the original tmp and discard the instruction.
* if the last conditional branch was turned into a jump, remove the
* fall-through edge to the next block.
*/
void TraceBuilder::reoptimize() {
FTRACE(5, "ReOptimize:vvvvvvvvvvvvvvvvvvvv\n");
SCOPE_EXIT { FTRACE(5, "ReOptimize:^^^^^^^^^^^^^^^^^^^^\n"); };
assert(m_curTrace->isMain());
assert(m_savedTraces.empty());
m_state.setEnableCse(RuntimeOption::EvalHHIRCse);
m_enableSimplification = RuntimeOption::EvalHHIRSimplification;
if (!m_state.enableCse() && !m_enableSimplification) return;
always_assert(!m_inReoptimize);
m_inReoptimize = true;
BlockList sortedBlocks = rpoSortCfg(m_unit);
auto const idoms = findDominators(m_unit, sortedBlocks);
m_state.clear();
auto& traceBlocks = m_curTrace->blocks();
BlockList blocks(traceBlocks.begin(), traceBlocks.end());
traceBlocks.clear();
for (auto* block : blocks) {
assert(block->trace() == m_curTrace);
FTRACE(5, "Block: {}\n", block->id());
assert(m_curTrace->isMain());
m_state.startBlock(block);
m_curTrace->push_back(block);
auto instructions = std::move(block->instrs());
assert(block->empty());
while (!instructions.empty()) {
auto *inst = &instructions.front();
instructions.pop_front();
m_state.setMarker(inst->marker());
// merging state looks at the current marker, and optimizeWork
// below may create new instructions. Use the marker from this
// instruction.
assert(inst->marker().valid());
setMarker(inst->marker());
auto const tmp = optimizeWork(inst, idoms); // Can generate new instrs!
if (!tmp) {
// Could not optimize; keep the old instruction
appendInstruction(inst, block);
m_state.update(inst);
continue;
}
SSATmp* dst = inst->dst();
if (dst->type() != Type::None && dst != tmp) {
// The result of optimization has a different destination than the inst.
// Generate a mov(tmp->dst) to get result into dst. If we get here then
// assume the last instruction in the block isn't a guard. If it was,
// we would have to insert the mov on the fall-through edge.
assert(block->empty() || !block->back().isBlockEnd());
IRInstruction* mov = m_unit.mov(dst, tmp, inst->marker());
appendInstruction(mov, block);
m_state.update(mov);
}
// Not re-adding inst; remove the inst->taken edge
if (inst->taken()) inst->setTaken(nullptr);
}
if (block->empty()) {
// If all the instructions in the block were optimized away, remove it
// from the trace.
auto it = traceBlocks.end();
--it;
assert(*it == block);
m_curTrace->unlink(it);
} else {
if (block->back().isTerminal()) {
// Could have converted a conditional branch to Jmp; clear next.
block->setNext(nullptr);
}
m_state.finishBlock(block);
}
}
}
/*
* reoptimize() runs a trace through a second pass of TraceBuilder
* optimizations, like this:
*
* reset state.
* move all blocks to a temporary list.
* compute immediate dominators.
* for each block in trace order:
* if we have a snapshot state for this block:
* clear cse entries that don't dominate this block.
* use snapshot state.
* move all instructions to a temporary list.
* for each instruction:
* optimizeWork - do CSE and simplify again
* if not simplified:
* append existing instruction and update state.
* else:
* if the instruction has a result, insert a mov from the
* simplified tmp to the original tmp and discard the instruction.
* if the last conditional branch was turned into a jump, remove the
* fall-through edge to the next block.
*/
void TraceBuilder::reoptimize() {
FTRACE(5, "ReOptimize:vvvvvvvvvvvvvvvvvvvv\n");
SCOPE_EXIT { FTRACE(5, "ReOptimize:^^^^^^^^^^^^^^^^^^^^\n"); };
assert(m_curTrace == m_mainTrace.get());
assert(m_savedTraces.empty());
assert(m_inlineSavedStates.empty());
m_enableCse = RuntimeOption::EvalHHIRCse;
m_enableSimplification = RuntimeOption::EvalHHIRSimplification;
if (!m_enableCse && !m_enableSimplification) return;
if (m_mainTrace->blocks().size() >
RuntimeOption::EvalHHIRSimplificationMaxBlocks) {
// TODO CSEHash::filter is very slow for large block sizes
// t2135219 should address that
return;
}
BlockList sortedBlocks = rpoSortCfg(m_mainTrace.get(), m_irFactory);
auto const idoms = findDominators(sortedBlocks);
clearTrackedState();
auto blocks = std::move(m_mainTrace->blocks());
assert(m_mainTrace->blocks().empty());
while (!blocks.empty()) {
Block* block = blocks.front();
blocks.pop_front();
assert(block->trace() == m_mainTrace.get());
FTRACE(5, "Block: {}\n", block->id());
m_mainTrace->push_back(block);
if (m_snapshots[block]) {
useState(block);
}
auto instructions = std::move(block->instrs());
assert(block->empty());
while (!instructions.empty()) {
auto *inst = &instructions.front();
instructions.pop_front();
// last attempt to elide ActRecs, if we still need the InlineFPAnchor
// it will be added back to the trace when we re-add instructions that
// rely on it
if (inst->op() == InlineFPAnchor) {
continue;
}
// merging state looks at the current marker, and optimizeWork
// below may create new instructions. Use the marker from this
// instruction.
assert(inst->marker().valid());
setMarker(inst->marker());
auto const tmp = optimizeWork(inst, idoms); // Can generate new instrs!
if (!tmp) {
// Could not optimize; keep the old instruction
appendInstruction(inst, block);
updateTrackedState(inst);
continue;
}
SSATmp* dst = inst->dst();
if (dst->type() != Type::None && dst != tmp) {
// The result of optimization has a different destination than the inst.
// Generate a mov(tmp->dst) to get result into dst. If we get here then
// assume the last instruction in the block isn't a guard. If it was,
// we would have to insert the mov on the fall-through edge.
assert(block->empty() || !block->back()->isBlockEnd());
IRInstruction* mov = m_irFactory.mov(dst, tmp, inst->marker());
appendInstruction(mov, block);
updateTrackedState(mov);
}
// Not re-adding inst; remove the inst->taken edge
if (inst->taken()) inst->setTaken(nullptr);
}
if (block->back()->isTerminal()) {
// Could have converted a conditional branch to Jmp; clear next.
block->setNext(nullptr);
} else {
// if the last instruction was a branch, we already saved state
// for the target in updateTrackedState(). Now save state for
// the fall-through path.
saveState(block->next());
}
}
}
/*
* reoptimize() runs a trace through a second pass of TraceBuilder
* optimizations, like this:
*
* reset state.
* move all blocks to a temporary list.
* compute immediate dominators.
* for each block in trace order:
* if we have a snapshot state for this block:
* clear cse entries that don't dominate this block.
* use snapshot state.
* move all instructions to a temporary list.
* for each instruction:
* optimizeWork - do CSE and simplify again
* if not simplified:
* append existing instruction and update state.
* else:
* if the instruction has a result, insert a mov from the
* simplified tmp to the original tmp and discard the instruction.
* if the last conditional branch was turned into a jump, remove the
* fall-through edge to the next block.
*/
void TraceBuilder::reoptimize() {
m_enableCse = RuntimeOption::EvalHHIRCse;
m_enableSimplification = RuntimeOption::EvalHHIRSimplification;
if (!m_enableCse && !m_enableSimplification) return;
if (m_trace->getBlocks().size() >
RuntimeOption::EvalHHIRSimplificationMaxBlocks) {
// TODO CSEHash::filter is very slow for large block sizes
// t2135219 should address that
return;
}
BlockList sortedBlocks = sortCfg(m_trace.get(), m_irFactory);
IdomVector idoms = findDominators(sortedBlocks);
clearTrackedState();
auto blocks = std::move(m_trace->getBlocks());
assert(m_trace->getBlocks().empty());
while (!blocks.empty()) {
Block* block = blocks.front();
blocks.pop_front();
assert(block->getTrace() == m_trace.get());
m_trace->push_back(block);
if (m_snapshots[block]) {
useState(block);
m_cseHash.filter(block, idoms);
}
auto instructions = std::move(block->getInstrs());
assert(block->empty());
while (!instructions.empty()) {
auto *inst = &instructions.front();
instructions.pop_front();
SSATmp* tmp = optimizeWork(inst); // Can generate new instrs!
if (!tmp) {
// Could not optimize; keep the old instruction
appendInstruction(inst, block);
updateTrackedState(inst);
continue;
}
SSATmp* dst = inst->getDst();
if (dst->type() != Type::None && dst != tmp) {
// The result of optimization has a different destination than the inst.
// Generate a mov(tmp->dst) to get result into dst. If we get here then
// assume the last instruction in the block isn't a guard. If it was,
// we would have to insert the mov on the fall-through edge.
assert(!block->back()->isBlockEnd());
IRInstruction* mov = m_irFactory.mov(dst, tmp);
appendInstruction(mov, block);
updateTrackedState(mov);
}
// Not re-adding inst; remove the inst->taken edge
if (inst->getTaken()) inst->setTaken(nullptr);
}
if (block->back()->isTerminal()) {
// Could have converted a conditional branch to Jmp; clear next.
block->setNext(nullptr);
} else {
// if the last instruction was a branch, we already saved state
// for the target in updateTrackedState(). Now save state for
// the fall-through path.
saveState(block->getNext());
}
}
}
/*
* reoptimize() runs a trace through a second pass of TraceBuilder
* optimizations, like this:
*
* reset state.
* move all blocks to a temporary list.
* compute immediate dominators.
* for each block in trace order:
* if we have a snapshot state for this block:
* clear cse entries that don't dominate this block.
* use snapshot state.
* move all instructions to a temporary list.
* for each instruction:
* optimizeWork - do CSE and simplify again
* if not simplified:
* append existing instruction and update state.
* else:
* if the instruction has a result, insert a mov from the
* simplified tmp to the original tmp and discard the instruction.
* if the last conditional branch was turned into a jump, remove the
* fall-through edge to the next block.
*/
void TraceBuilder::reoptimize() {
FTRACE(5, "ReOptimize:vvvvvvvvvvvvvvvvvvvv\n");
SCOPE_EXIT { FTRACE(5, "ReOptimize:^^^^^^^^^^^^^^^^^^^^\n"); };
assert(m_savedBlocks.empty());
assert(!m_curWhere);
m_state.setEnableCse(RuntimeOption::EvalHHIRCse);
m_enableSimplification = RuntimeOption::EvalHHIRSimplification;
if (!m_state.enableCse() && !m_enableSimplification) return;
setConstrainGuards(false);
BlockList sortedBlocks = rpoSortCfg(m_unit);
auto const idoms = findDominators(m_unit, sortedBlocks);
m_state.clear();
for (auto* block : rpoSortCfg(m_unit)) {
FTRACE(5, "Block: {}\n", block->id());
m_state.startBlock(block);
m_curBlock = block;
auto instructions = std::move(block->instrs());
assert(block->empty());
while (!instructions.empty()) {
auto *inst = &instructions.front();
instructions.pop_front();
// merging state looks at the current marker, and optimizeWork
// below may create new instructions. Use the marker from this
// instruction.
assert(inst->marker().valid());
setMarker(inst->marker());
auto const tmp = optimizeWork(inst, idoms); // Can generate new instrs!
if (!tmp) {
// Could not optimize; keep the old instruction
appendInstruction(inst);
continue;
}
SSATmp* dst = inst->dst();
if (dst->type() != Type::None && dst != tmp) {
// The result of optimization has a different destination than the inst.
// Generate a mov(tmp->dst) to get result into dst. If we get here then
// assume the last instruction in the block isn't a guard. If it was,
// we would have to insert the mov on the fall-through edge.
assert(block->empty() || !block->back().isBlockEnd());
IRInstruction* mov = m_unit.mov(dst, tmp, inst->marker());
appendInstruction(mov);
}
if (inst->isBlockEnd()) {
// Not re-adding inst; replace it with a jump to the next block.
auto next = inst->next();
appendInstruction(m_unit.gen(Jmp, inst->marker(), next));
inst->setTaken(nullptr);
inst->setNext(nullptr);
}
}
assert(!block->empty());
m_state.finishBlock(block);
}
}
