Add trace accessor methods, implement primitive if-conversion heuristic.
Compare the critical paths of the two traces through an if-conversion
candidate. If the difference is larger than the branch brediction
penalty, reject the if-conversion. If would never pay.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@161433 91177308-0d34-0410-b5e6-96231b3b80d8
diff --git a/lib/CodeGen/EarlyIfConversion.cpp b/lib/CodeGen/EarlyIfConversion.cpp
index e49b9d3..8adaa6d 100644
--- a/lib/CodeGen/EarlyIfConversion.cpp
+++ b/lib/CodeGen/EarlyIfConversion.cpp
@@ -601,10 +601,24 @@
bool EarlyIfConverter::shouldConvertIf() {
if (!MinInstr)
MinInstr = Traces->getEnsemble(MachineTraceMetrics::TS_MinInstrCount);
- DEBUG({
- dbgs() << MinInstr->getTrace(IfConv.Head);
- MinInstr->print(dbgs());
- });
+
+ // MCSchedModel doesn't yet provide a misprediction penalty.
+ unsigned MispredictPenalty = 10;
+
+ // Compare the critical path through TBB and FBB. If the difference is
+ // greater than the branch misprediction penalty, it would never pay to
+ // if-convert. The triangle/diamond topology guarantees that these traces
+ // have the same head and tail, so they can be compared.
+ MachineTraceMetrics::Trace TBBTrace = MinInstr->getTrace(IfConv.TBB);
+ MachineTraceMetrics::Trace FBBTrace = MinInstr->getTrace(IfConv.FBB);
+ DEBUG(dbgs() << "TBB: " << TBBTrace << "FBB: " << FBBTrace);
+ unsigned TBBCrit = TBBTrace.getCriticalPath();
+ unsigned FBBCrit = FBBTrace.getCriticalPath();
+ unsigned ExtraCrit = TBBCrit > FBBCrit ? TBBCrit-FBBCrit : FBBCrit-TBBCrit;
+ if (ExtraCrit >= MispredictPenalty) {
+ DEBUG(dbgs() << "Critical path difference too large.\n");
+ return false;
+ }
return true;
}
diff --git a/lib/CodeGen/MachineTraceMetrics.cpp b/lib/CodeGen/MachineTraceMetrics.cpp
index 1c0894e..09b6e9c 100644
--- a/lib/CodeGen/MachineTraceMetrics.cpp
+++ b/lib/CodeGen/MachineTraceMetrics.cpp
@@ -14,6 +14,7 @@
#include "llvm/CodeGen/MachineLoopInfo.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/CodeGen/Passes.h"
+#include "llvm/MC/MCInstrItineraries.h"
#include "llvm/Target/TargetInstrInfo.h"
#include "llvm/Target/TargetRegisterInfo.h"
#include "llvm/Support/Debug.h"
@@ -1017,6 +1018,29 @@
return Trace(*this, BlockInfo[MBB->getNumber()]);
}
+unsigned
+MachineTraceMetrics::Trace::getInstrSlack(const MachineInstr *MI) const {
+ assert(MI && "Not an instruction.");
+ assert(getBlockNum() == unsigned(MI->getParent()->getNumber()) &&
+ "MI must be in the trace center block");
+ InstrCycles Cyc = getInstrCycles(MI);
+ return getCriticalPath() - (Cyc.Depth + Cyc.Height);
+}
+
+unsigned MachineTraceMetrics::Trace::getResourceDepth(bool Bottom) const {
+ // For now, we compute the resource depth from instruction count / issue
+ // width. Eventually, we should compute resource depth per functional unit
+ // and return the max.
+ unsigned Instrs = TBI.InstrDepth;
+ if (Bottom)
+ Instrs += TE.MTM.BlockInfo[getBlockNum()].InstrCount;
+ if (const MCSchedModel *Model = TE.MTM.ItinData->SchedModel)
+ if (Model->IssueWidth != 0)
+ return Instrs / Model->IssueWidth;
+ // Assume issue width 1 without a schedule model.
+ return Instrs;
+}
+
void MachineTraceMetrics::Ensemble::print(raw_ostream &OS) const {
OS << getName() << " ensemble:\n";
for (unsigned i = 0, e = BlockInfo.size(); i != e; ++i) {
diff --git a/lib/CodeGen/MachineTraceMetrics.h b/lib/CodeGen/MachineTraceMetrics.h
index 33c2d95..626b28b 100644
--- a/lib/CodeGen/MachineTraceMetrics.h
+++ b/lib/CodeGen/MachineTraceMetrics.h
@@ -188,23 +188,6 @@
void print(raw_ostream&) const;
};
- /// A trace represents a plausible sequence of executed basic blocks that
- /// passes through the current basic block one. The Trace class serves as a
- /// handle to internal cached data structures.
- class Trace {
- Ensemble &TE;
- TraceBlockInfo &TBI;
-
- public:
- explicit Trace(Ensemble &te, TraceBlockInfo &tbi) : TE(te), TBI(tbi) {}
- void print(raw_ostream&) const;
-
- /// Compute the total number of instructions in the trace.
- unsigned getInstrCount() const {
- return TBI.InstrDepth + TBI.InstrHeight;
- }
- };
-
/// InstrCycles represents the cycle height and depth of an instruction in a
/// trace.
struct InstrCycles {
@@ -218,6 +201,48 @@
unsigned Height;
};
+ /// A trace represents a plausible sequence of executed basic blocks that
+ /// passes through the current basic block one. The Trace class serves as a
+ /// handle to internal cached data structures.
+ class Trace {
+ Ensemble &TE;
+ TraceBlockInfo &TBI;
+
+ unsigned getBlockNum() const { return &TBI - &TE.BlockInfo[0]; }
+
+ public:
+ explicit Trace(Ensemble &te, TraceBlockInfo &tbi) : TE(te), TBI(tbi) {}
+ void print(raw_ostream&) const;
+
+ /// Compute the total number of instructions in the trace.
+ unsigned getInstrCount() const {
+ return TBI.InstrDepth + TBI.InstrHeight;
+ }
+
+ /// Return the resource dpeth of the top/bottom of the trace center block.
+ /// This is the number of cycles required to execute all instructions from
+ /// the trace head to the trace center block. The resource depth only
+ /// considers execution resources, it ignores data dependencies.
+ /// When Bottom is set, instructions in the trace center block are included.
+ unsigned getResourceDepth(bool Bottom) const;
+
+ /// Return the length of the (data dependency) critical path through the
+ /// trace.
+ unsigned getCriticalPath() const { return TBI.CriticalPath; }
+
+ /// Return the depth and height of MI. The depth is only valid for
+ /// instructions in or above the trace center block. The height is only
+ /// valid for instructions in or below the trace center block.
+ InstrCycles getInstrCycles(const MachineInstr *MI) const {
+ return TE.Cycles.lookup(MI);
+ }
+
+ /// Return the slack of MI. This is the number of cycles MI can be delayed
+ /// before the critical path becomes longer.
+ /// MI must be an instruction in the trace center block.
+ unsigned getInstrSlack(const MachineInstr *MI) const;
+ };
+
/// A trace ensemble is a collection of traces selected using the same
/// strategy, for example 'minimum resource height'. There is one trace for
/// every block in the function.