llvm/lib/CodeGen/InstrSched/SchedPriorities.cpp

284 lines
9.3 KiB
C++
Raw Normal View History

//===-- SchedPriorities.h - Encapsulate scheduling heuristics -------------===//
//
// The LLVM Compiler Infrastructure
//
// This file was developed by the LLVM research group and is distributed under
// the University of Illinois Open Source License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
// Strategy:
// Priority ordering rules:
// (1) Max delay, which is the order of the heap S.candsAsHeap.
// (2) Instruction that frees up a register.
// (3) Instruction that has the maximum number of dependent instructions.
// Note that rules 2 and 3 are only used if issue conflicts prevent
// choosing a higher priority instruction by rule 1.
//
//===----------------------------------------------------------------------===//
#include "SchedPriorities.h"
#include "llvm/CodeGen/FunctionLiveVarInfo.h"
#include "llvm/CodeGen/MachineBasicBlock.h"
#include "llvm/Support/CFG.h"
#include "Support/PostOrderIterator.h"
namespace llvm {
std::ostream &operator<<(std::ostream &os, const NodeDelayPair* nd) {
return os << "Delay for node " << nd->node->getNodeId()
<< " = " << (long)nd->delay << "\n";
}
SchedPriorities::SchedPriorities(const Function *, const SchedGraph *G,
FunctionLiveVarInfo &LVI)
: curTime(0), graph(G), methodLiveVarInfo(LVI),
nodeDelayVec(G->getNumNodes(), INVALID_LATENCY), // make errors obvious
earliestReadyTimeForNode(G->getNumNodes(), 0),
earliestReadyTime(0),
nextToTry(candsAsHeap.begin())
{
computeDelays(graph);
}
void
SchedPriorities::initialize() {
initializeReadyHeap(graph);
}
void
SchedPriorities::computeDelays(const SchedGraph* graph) {
po_iterator<const SchedGraph*> poIter = po_begin(graph), poEnd =po_end(graph);
for ( ; poIter != poEnd; ++poIter) {
const SchedGraphNode* node = *poIter;
cycles_t nodeDelay;
if (node->beginOutEdges() == node->endOutEdges())
nodeDelay = node->getLatency();
else {
// Iterate over the out-edges of the node to compute delay
nodeDelay = 0;
for (SchedGraphNode::const_iterator E=node->beginOutEdges();
E != node->endOutEdges(); ++E) {
cycles_t sinkDelay = getNodeDelay((SchedGraphNode*)(*E)->getSink());
nodeDelay = std::max(nodeDelay, sinkDelay + (*E)->getMinDelay());
}
}
getNodeDelayRef(node) = nodeDelay;
}
}
void
SchedPriorities::initializeReadyHeap(const SchedGraph* graph) {
const SchedGraphNode* graphRoot = (const SchedGraphNode*)graph->getRoot();
assert(graphRoot->getMachineInstr() == NULL && "Expect dummy root");
// Insert immediate successors of dummy root, which are the actual roots
sg_succ_const_iterator SEnd = succ_end(graphRoot);
for (sg_succ_const_iterator S = succ_begin(graphRoot); S != SEnd; ++S)
this->insertReady(*S);
#undef TEST_HEAP_CONVERSION
#ifdef TEST_HEAP_CONVERSION
std::cerr << "Before heap conversion:\n";
copy(candsAsHeap.begin(), candsAsHeap.end(),
ostream_iterator<NodeDelayPair*>(std::cerr,"\n"));
#endif
candsAsHeap.makeHeap();
nextToTry = candsAsHeap.begin();
#ifdef TEST_HEAP_CONVERSION
std::cerr << "After heap conversion:\n";
copy(candsAsHeap.begin(), candsAsHeap.end(),
ostream_iterator<NodeDelayPair*>(std::cerr,"\n"));
#endif
}
void
SchedPriorities::insertReady(const SchedGraphNode* node) {
candsAsHeap.insert(node, nodeDelayVec[node->getNodeId()]);
candsAsSet.insert(node);
mcands.clear(); // ensure reset choices is called before any more choices
earliestReadyTime = std::min(earliestReadyTime,
getEarliestReadyTimeForNode(node));
if (SchedDebugLevel >= Sched_PrintSchedTrace) {
std::cerr << " Node " << node->getNodeId() << " will be ready in Cycle "
<< getEarliestReadyTimeForNode(node) << "; "
<< " Delay = " <<(long)getNodeDelay(node) << "; Instruction: \n"
<< " " << *node->getMachineInstr() << "\n";
}
}
void
SchedPriorities::issuedReadyNodeAt(cycles_t curTime,
const SchedGraphNode* node) {
candsAsHeap.removeNode(node);
candsAsSet.erase(node);
mcands.clear(); // ensure reset choices is called before any more choices
if (earliestReadyTime == getEarliestReadyTimeForNode(node)) {
// earliestReadyTime may have been due to this node, so recompute it
earliestReadyTime = HUGE_LATENCY;
for (NodeHeap::const_iterator I=candsAsHeap.begin();
I != candsAsHeap.end(); ++I)
if (candsAsHeap.getNode(I)) {
earliestReadyTime =
std::min(earliestReadyTime,
getEarliestReadyTimeForNode(candsAsHeap.getNode(I)));
}
}
// Now update ready times for successors
for (SchedGraphNode::const_iterator E=node->beginOutEdges();
E != node->endOutEdges(); ++E) {
cycles_t& etime =
getEarliestReadyTimeForNodeRef((SchedGraphNode*)(*E)->getSink());
etime = std::max(etime, curTime + (*E)->getMinDelay());
}
}
//----------------------------------------------------------------------
// Priority ordering rules:
// (1) Max delay, which is the order of the heap S.candsAsHeap.
// (2) Instruction that frees up a register.
// (3) Instruction that has the maximum number of dependent instructions.
// Note that rules 2 and 3 are only used if issue conflicts prevent
// choosing a higher priority instruction by rule 1.
//----------------------------------------------------------------------
inline int
SchedPriorities::chooseByRule1(std::vector<candIndex>& mcands) {
return (mcands.size() == 1)? 0 // only one choice exists so take it
: -1; // -1 indicates multiple choices
}
inline int
SchedPriorities::chooseByRule2(std::vector<candIndex>& mcands) {
assert(mcands.size() >= 1 && "Should have at least one candidate here.");
for (unsigned i=0, N = mcands.size(); i < N; i++)
if (instructionHasLastUse(methodLiveVarInfo,
candsAsHeap.getNode(mcands[i])))
return i;
return -1;
}
inline int
SchedPriorities::chooseByRule3(std::vector<candIndex>& mcands) {
assert(mcands.size() >= 1 && "Should have at least one candidate here.");
int maxUses = candsAsHeap.getNode(mcands[0])->getNumOutEdges();
int indexWithMaxUses = 0;
for (unsigned i=1, N = mcands.size(); i < N; i++) {
int numUses = candsAsHeap.getNode(mcands[i])->getNumOutEdges();
if (numUses > maxUses) {
maxUses = numUses;
indexWithMaxUses = i;
}
}
return indexWithMaxUses;
}
const SchedGraphNode*
SchedPriorities::getNextHighest(const SchedulingManager& S,
cycles_t curTime) {
int nextIdx = -1;
const SchedGraphNode* nextChoice = NULL;
if (mcands.size() == 0)
findSetWithMaxDelay(mcands, S);
while (nextIdx < 0 && mcands.size() > 0) {
nextIdx = chooseByRule1(mcands); // rule 1
if (nextIdx == -1)
nextIdx = chooseByRule2(mcands); // rule 2
if (nextIdx == -1)
nextIdx = chooseByRule3(mcands); // rule 3
if (nextIdx == -1)
nextIdx = 0; // default to first choice by delays
// We have found the next best candidate. Check if it ready in
// the current cycle, and if it is feasible.
// If not, remove it from mcands and continue. Refill mcands if
// it becomes empty.
nextChoice = candsAsHeap.getNode(mcands[nextIdx]);
if (getEarliestReadyTimeForNode(nextChoice) > curTime
|| ! instrIsFeasible(S, nextChoice->getMachineInstr()->getOpcode()))
{
mcands.erase(mcands.begin() + nextIdx);
nextIdx = -1;
if (mcands.size() == 0)
findSetWithMaxDelay(mcands, S);
}
}
if (nextIdx >= 0) {
mcands.erase(mcands.begin() + nextIdx);
return nextChoice;
} else
return NULL;
}
void
SchedPriorities::findSetWithMaxDelay(std::vector<candIndex>& mcands,
const SchedulingManager& S)
{
if (mcands.size() == 0 && nextToTry != candsAsHeap.end())
{ // out of choices at current maximum delay;
// put nodes with next highest delay in mcands
candIndex next = nextToTry;
cycles_t maxDelay = candsAsHeap.getDelay(next);
for (; next != candsAsHeap.end()
&& candsAsHeap.getDelay(next) == maxDelay; ++next)
mcands.push_back(next);
nextToTry = next;
if (SchedDebugLevel >= Sched_PrintSchedTrace) {
std::cerr << " Cycle " << (long)getTime() << ": "
<< "Next highest delay = " << (long)maxDelay << " : "
<< mcands.size() << " Nodes with this delay: ";
for (unsigned i=0; i < mcands.size(); i++)
std::cerr << candsAsHeap.getNode(mcands[i])->getNodeId() << ", ";
std::cerr << "\n";
}
}
}
bool
SchedPriorities::instructionHasLastUse(FunctionLiveVarInfo &LVI,
const SchedGraphNode* graphNode) {
const MachineInstr *MI = graphNode->getMachineInstr();
hash_map<const MachineInstr*, bool>::const_iterator
ui = lastUseMap.find(MI);
if (ui != lastUseMap.end())
return ui->second;
// else check if instruction is a last use and save it in the hash_map
bool hasLastUse = false;
const BasicBlock* bb = graphNode->getMachineBasicBlock().getBasicBlock();
const ValueSet &LVs = LVI.getLiveVarSetBeforeMInst(MI, bb);
for (MachineInstr::const_val_op_iterator OI = MI->begin(), OE = MI->end();
OI != OE; ++OI)
if (!LVs.count(*OI)) {
hasLastUse = true;
break;
}
return lastUseMap[MI] = hasLastUse;
}
} // End llvm namespace