darling-JavaScriptCore/dfg/DFGCFAPhase.cpp

285 lines
12 KiB
C++

/*
* Copyright (C) 2011-2018 Apple Inc. All rights reserved.
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions
* are met:
* 1. Redistributions of source code must retain the above copyright
* notice, this list of conditions and the following disclaimer.
* 2. Redistributions in binary form must reproduce the above copyright
* notice, this list of conditions and the following disclaimer in the
* documentation and/or other materials provided with the distribution.
*
* THIS SOFTWARE IS PROVIDED BY APPLE INC. ``AS IS'' AND ANY
* EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
* PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL APPLE INC. OR
* CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
* EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
* PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
* PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
* OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*/
#include "config.h"
#include "DFGCFAPhase.h"
#if ENABLE(DFG_JIT)
#include "DFGAbstractInterpreterInlines.h"
#include "DFGBlockSet.h"
#include "DFGClobberSet.h"
#include "DFGClobberize.h"
#include "DFGGraph.h"
#include "DFGInPlaceAbstractState.h"
#include "DFGPhase.h"
#include "DFGSafeToExecute.h"
#include "OperandsInlines.h"
#include "JSCInlines.h"
namespace JSC { namespace DFG {
class CFAPhase : public Phase {
public:
CFAPhase(Graph& graph)
: Phase(graph, "control flow analysis")
, m_state(graph)
, m_interpreter(graph, m_state)
, m_verbose(Options::verboseCFA())
{
}
bool run()
{
ASSERT(m_graph.m_form == ThreadedCPS || m_graph.m_form == SSA);
ASSERT(m_graph.m_unificationState == GloballyUnified);
ASSERT(m_graph.m_refCountState == EverythingIsLive);
m_count = 0;
if (m_verbose && !shouldDumpGraphAtEachPhase(m_graph.m_plan.mode())) {
dataLog("Graph before CFA:\n");
m_graph.dump();
}
// This implements a pseudo-worklist-based forward CFA, except that the visit order
// of blocks is the bytecode program order (which is nearly topological), and
// instead of a worklist we just walk all basic blocks checking if cfaShouldRevisit
// is set to true. This is likely to balance the efficiency properties of both
// worklist-based and forward fixpoint-based approaches. Like a worklist-based
// approach, it won't visit code if it's meaningless to do so (nothing changed at
// the head of the block or the predecessors have not been visited). Like a forward
// fixpoint-based approach, it has a high probability of only visiting a block
// after all predecessors have been visited. Only loops will cause this analysis to
// revisit blocks, and the amount of revisiting is proportional to loop depth.
m_state.initialize();
if (m_graph.m_form != SSA) {
if (m_verbose)
dataLog(" Widening state at OSR entry block.\n");
// Widen the abstract values at the block that serves as the must-handle OSR entry.
for (BlockIndex blockIndex = m_graph.numBlocks(); blockIndex--;) {
BasicBlock* block = m_graph.block(blockIndex);
if (!block)
continue;
if (!block->isOSRTarget)
continue;
if (block->bytecodeBegin != m_graph.m_plan.osrEntryBytecodeIndex())
continue;
// We record that the block needs some OSR stuff, but we don't do that yet. We want to
// handle OSR entry data at the right time in order to get the best compile times. If we
// simply injected OSR data right now, then we'd potentially cause a loop body to be
// interpreted with just the constants we feed it, which is more expensive than if we
// interpreted it with non-constant values. If we always injected this data after the
// main pass of CFA ran, then we would potentially spend a bunch of time rerunning CFA
// after convergence. So, we try very hard to inject OSR data for a block when we first
// naturally come to see it - see the m_blocksWithOSR check in performBlockCFA(). This
// way, we:
//
// - Reduce the likelihood of interpreting the block with constants, since we will inject
// the OSR entry constants on top of whatever abstract values we got for that block on
// the first pass. The mix of those two things is likely to not be constant.
//
// - Reduce the total number of CFA reexecutions since we inject the OSR data as part of
// the normal flow of CFA instead of having to do a second fixpoint. We may still have
// to do a second fixpoint if we don't even reach the OSR entry block during the main
// run of CFA, but in that case at least we're not being redundant.
m_blocksWithOSR.add(block);
}
}
do {
m_changed = false;
performForwardCFA();
} while (m_changed);
if (m_graph.m_form != SSA) {
for (BlockIndex blockIndex = m_graph.numBlocks(); blockIndex--;) {
BasicBlock* block = m_graph.block(blockIndex);
if (!block)
continue;
if (m_blocksWithOSR.remove(block))
m_changed |= injectOSR(block);
}
while (m_changed) {
m_changed = false;
performForwardCFA();
}
// Make sure we record the intersection of all proofs that we ever allowed the
// compiler to rely upon.
for (BlockIndex blockIndex = m_graph.numBlocks(); blockIndex--;) {
BasicBlock* block = m_graph.block(blockIndex);
if (!block)
continue;
block->intersectionOfCFAHasVisited &= block->cfaHasVisited;
for (unsigned i = block->intersectionOfPastValuesAtHead.size(); i--;) {
AbstractValue value = block->valuesAtHead[i];
// We need to guarantee that when we do an OSR entry, we validate the incoming
// value as if it could be live past an invalidation point. Otherwise, we may
// OSR enter with a value with the wrong structure, and an InvalidationPoint's
// promise of filtering the structure set of certain values is no longer upheld.
value.m_structure.observeInvalidationPoint();
block->intersectionOfPastValuesAtHead[i].filter(value);
}
}
}
return true;
}
private:
bool injectOSR(BasicBlock* block)
{
if (m_verbose)
dataLog(" Found must-handle block: ", *block, "\n");
// This merges snapshot of stack values while CFA phase want to have proven types and values. This is somewhat tricky.
// But this is OK as long as DFG OSR entry validates the inputs with *proven* AbstracValue values. And it turns out that this
// type widening is critical to navier-stokes. Without it, navier-stokes has more strict constraint on OSR entry and
// fails OSR entry repeatedly.
bool changed = false;
const Operands<Optional<JSValue>>& mustHandleValues = m_graph.m_plan.mustHandleValues();
for (size_t i = mustHandleValues.size(); i--;) {
Operand operand = mustHandleValues.operandForIndex(i);
Optional<JSValue> value = mustHandleValues[i];
if (!value) {
if (m_verbose)
dataLog(" Not live in bytecode: ", operand, "\n");
continue;
}
Node* node = block->variablesAtHead.operand(operand);
if (!node) {
if (m_verbose)
dataLog(" Not live: ", operand, "\n");
continue;
}
if (m_verbose)
dataLog(" Widening ", operand, " with ", value.value(), "\n");
AbstractValue& target = block->valuesAtHead.operand(operand);
changed |= target.mergeOSREntryValue(m_graph, value.value(), node->variableAccessData(), node);
}
if (changed || !block->cfaHasVisited) {
block->cfaShouldRevisit = true;
return true;
}
return false;
}
void performBlockCFA(BasicBlock* block)
{
if (!block)
return;
if (!block->cfaShouldRevisit)
return;
if (m_verbose)
dataLog(" Block ", *block, ":\n");
if (m_blocksWithOSR.remove(block))
injectOSR(block);
m_state.beginBasicBlock(block);
if (m_verbose) {
dataLog(" head vars: ", block->valuesAtHead, "\n");
if (m_graph.m_form == SSA)
dataLog(" head regs: ", nodeValuePairListDump(block->ssa->valuesAtHead), "\n");
}
for (unsigned i = 0; i < block->size(); ++i) {
Node* node = block->at(i);
if (m_verbose) {
dataLogF(" %s @%u: ", Graph::opName(node->op()), node->index());
if (!safeToExecute(m_state, m_graph, node))
dataLog("(UNSAFE) ");
dataLog(m_state.variablesForDebugging(), " ", m_interpreter);
dataLogF("\n");
}
if (!m_interpreter.execute(i)) {
if (m_verbose)
dataLogF(" Expect OSR exit.\n");
break;
}
if (ASSERT_ENABLED
&& m_state.didClobberOrFolded() != writesOverlap(m_graph, node, JSCell_structureID))
DFG_CRASH(m_graph, node, toCString("AI-clobberize disagreement; AI says ", m_state.clobberState(), " while clobberize says ", writeSet(m_graph, node)).data());
}
if (m_verbose) {
dataLogF(" tail regs: ");
m_interpreter.dump(WTF::dataFile());
dataLogF("\n");
}
m_changed |= m_state.endBasicBlock();
if (m_verbose) {
dataLog(" tail vars: ", block->valuesAtTail, "\n");
if (m_graph.m_form == SSA)
dataLog(" head regs: ", nodeValuePairListDump(block->ssa->valuesAtTail), "\n");
}
}
void performForwardCFA()
{
++m_count;
if (m_verbose)
dataLogF("CFA [%u]\n", m_count);
for (BlockIndex blockIndex = 0; blockIndex < m_graph.numBlocks(); ++blockIndex)
performBlockCFA(m_graph.block(blockIndex));
}
private:
InPlaceAbstractState m_state;
AbstractInterpreter<InPlaceAbstractState> m_interpreter;
BlockSet m_blocksWithOSR;
const bool m_verbose;
bool m_changed;
unsigned m_count;
};
bool performCFA(Graph& graph)
{
return runPhase<CFAPhase>(graph);
}
} } // namespace JSC::DFG
#endif // ENABLE(DFG_JIT)