1 //===- lib/CodeGen/MachineTraceMetrics.h - Super-scalar metrics -*- C++ -*-===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines the interface for the MachineTraceMetrics analysis pass
11 // that estimates CPU resource usage and critical data dependency paths through
12 // preferred traces. This is useful for super-scalar CPUs where execution speed
13 // can be limited both by data dependencies and by limited execution resources.
15 // Out-of-order CPUs will often be executing instructions from multiple basic
16 // blocks at the same time. This makes it difficult to estimate the resource
17 // usage accurately in a single basic block. Resources can be estimated better
18 // by looking at a trace through the current basic block.
20 // For every block, the MachineTraceMetrics pass will pick a preferred trace
21 // that passes through the block. The trace is chosen based on loop structure,
22 // branch probabilities, and resource usage. The intention is to pick likely
23 // traces that would be the most affected by code transformations.
25 // It is expensive to compute a full arbitrary trace for every block, so to
26 // save some computations, traces are chosen to be convergent. This means that
27 // if the traces through basic blocks A and B ever cross when moving away from
28 // A and B, they never diverge again. This applies in both directions - If the
29 // traces meet above A and B, they won't diverge when going further back.
31 // Traces tend to align with loops. The trace through a block in an inner loop
32 // will begin at the loop entry block and end at a back edge. If there are
33 // nested loops, the trace may begin and end at those instead.
35 // For each trace, we compute the critical path length, which is the number of
36 // cycles required to execute the trace when execution is limited by data
37 // dependencies only. We also compute the resource height, which is the number
38 // of cycles required to execute all instructions in the trace when ignoring
41 // Every instruction in the current block has a slack - the number of cycles
42 // execution of the instruction can be delayed without extending the critical
45 //===----------------------------------------------------------------------===//
47 #ifndef LLVM_CODEGEN_MACHINE_TRACE_METRICS_H
48 #define LLVM_CODEGEN_MACHINE_TRACE_METRICS_H
50 #include "llvm/ADT/DenseMap.h"
51 #include "llvm/ADT/SmallVector.h"
52 #include "llvm/CodeGen/MachineFunctionPass.h"
56 class InstrItineraryData;
57 class MachineBasicBlock;
60 class MachineLoopInfo;
61 class MachineRegisterInfo;
62 class TargetInstrInfo;
63 class TargetRegisterInfo;
66 class MachineTraceMetrics : public MachineFunctionPass {
67 const MachineFunction *MF;
68 const TargetInstrInfo *TII;
69 const TargetRegisterInfo *TRI;
70 const InstrItineraryData *ItinData;
71 const MachineRegisterInfo *MRI;
72 const MachineLoopInfo *Loops;
78 MachineTraceMetrics();
79 void getAnalysisUsage(AnalysisUsage&) const;
80 bool runOnMachineFunction(MachineFunction&);
82 void verifyAnalysis() const;
84 friend class Ensemble;
87 /// Per-basic block information that doesn't depend on the trace through the
89 struct FixedBlockInfo {
90 /// The number of non-trivial instructions in the block.
91 /// Doesn't count PHI and COPY instructions that are likely to be removed.
94 /// True when the block contains calls.
97 FixedBlockInfo() : InstrCount(~0u), HasCalls(false) {}
99 /// Returns true when resource information for this block has been computed.
100 bool hasResources() const { return InstrCount != ~0u; }
102 /// Invalidate resource information.
103 void invalidate() { InstrCount = ~0u; }
106 /// Get the fixed resource information about MBB. Compute it on demand.
107 const FixedBlockInfo *getResources(const MachineBasicBlock*);
109 /// Per-basic block information that relates to a specific trace through the
110 /// block. Convergent traces means that only one of these is required per
111 /// block in a trace ensemble.
112 struct TraceBlockInfo {
113 /// Trace predecessor, or NULL for the first block in the trace.
114 /// Valid when hasValidDepth().
115 const MachineBasicBlock *Pred;
117 /// Trace successor, or NULL for the last block in the trace.
118 /// Valid when hasValidHeight().
119 const MachineBasicBlock *Succ;
121 /// The block number of the head of the trace. (When hasValidDepth()).
124 /// The block number of the tail of the trace. (When hasValidHeight()).
127 /// Accumulated number of instructions in the trace above this block.
128 /// Does not include instructions in this block.
131 /// Accumulated number of instructions in the trace below this block.
132 /// Includes instructions in this block.
133 unsigned InstrHeight;
137 InstrDepth(~0u), InstrHeight(~0u),
138 HasValidInstrDepths(false), HasValidInstrHeights(false) {}
140 /// Returns true if the depth resources have been computed from the trace
141 /// above this block.
142 bool hasValidDepth() const { return InstrDepth != ~0u; }
144 /// Returns true if the height resources have been computed from the trace
145 /// below this block.
146 bool hasValidHeight() const { return InstrHeight != ~0u; }
148 /// Invalidate depth resources when some block above this one has changed.
149 void invalidateDepth() { InstrDepth = ~0u; HasValidInstrDepths = false; }
151 /// Invalidate height resources when a block below this one has changed.
152 void invalidateHeight() { InstrHeight = ~0u; HasValidInstrHeights = false; }
154 // Data-dependency-related information. Per-instruction depth and height
155 // are computed from data dependencies in the current trace, using
158 /// Instruction depths have been computed. This implies hasValidDepth().
159 bool HasValidInstrDepths;
161 /// Instruction heights have been computed. This implies hasValidHeight().
162 bool HasValidInstrHeights;
164 void print(raw_ostream&) const;
167 /// A trace represents a plausible sequence of executed basic blocks that
168 /// passes through the current basic block one. The Trace class serves as a
169 /// handle to internal cached data structures.
175 explicit Trace(Ensemble &te, TraceBlockInfo &tbi) : TE(te), TBI(tbi) {}
176 void print(raw_ostream&) const;
178 /// Compute the total number of instructions in the trace.
179 unsigned getInstrCount() const {
180 return TBI.InstrDepth + TBI.InstrHeight;
184 /// InstrCycles represents the cycle height and depth of an instruction in a
187 /// Earliest issue cycle as determined by data dependencies and instruction
188 /// latencies from the beginning of the trace. Data dependencies from
189 /// before the trace are not included.
192 /// Minimum number of cycles from this instruction is issued to the of the
193 /// trace, as determined by data dependencies and instruction latencies.
197 /// A trace ensemble is a collection of traces selected using the same
198 /// strategy, for example 'minimum resource height'. There is one trace for
199 /// every block in the function.
201 SmallVector<TraceBlockInfo, 4> BlockInfo;
202 DenseMap<const MachineInstr*, InstrCycles> Cycles;
205 void computeTrace(const MachineBasicBlock*);
206 void computeDepthResources(const MachineBasicBlock*);
207 void computeHeightResources(const MachineBasicBlock*);
208 void computeInstrDepths(const MachineBasicBlock*);
209 void computeInstrHeights(const MachineBasicBlock*);
212 MachineTraceMetrics &MTM;
213 virtual const MachineBasicBlock *pickTracePred(const MachineBasicBlock*) =0;
214 virtual const MachineBasicBlock *pickTraceSucc(const MachineBasicBlock*) =0;
215 explicit Ensemble(MachineTraceMetrics*);
216 const MachineLoop *getLoopFor(const MachineBasicBlock*) const;
217 const TraceBlockInfo *getDepthResources(const MachineBasicBlock*) const;
218 const TraceBlockInfo *getHeightResources(const MachineBasicBlock*) const;
222 virtual const char *getName() const =0;
223 void print(raw_ostream&) const;
224 void invalidate(const MachineBasicBlock *MBB);
227 /// Get the trace that passes through MBB.
228 /// The trace is computed on demand.
229 Trace getTrace(const MachineBasicBlock *MBB);
232 /// Strategies for selecting traces.
234 /// Select the trace through a block that has the fewest instructions.
240 /// Get the trace ensemble representing the given trace selection strategy.
241 /// The returned Ensemble object is owned by the MachineTraceMetrics analysis,
242 /// and valid for the lifetime of the analysis pass.
243 Ensemble *getEnsemble(Strategy);
245 /// Invalidate cached information about MBB. This must be called *before* MBB
246 /// is erased, or the CFG is otherwise changed.
248 /// This invalidates per-block information about resource usage for MBB only,
249 /// and it invalidates per-trace information for any trace that passes
252 /// Call Ensemble::getTrace() again to update any trace handles.
253 void invalidate(const MachineBasicBlock *MBB);
256 // One entry per basic block, indexed by block number.
257 SmallVector<FixedBlockInfo, 4> BlockInfo;
259 // One ensemble per strategy.
260 Ensemble* Ensembles[TS_NumStrategies];
263 inline raw_ostream &operator<<(raw_ostream &OS,
264 const MachineTraceMetrics::Trace &Tr) {
269 inline raw_ostream &operator<<(raw_ostream &OS,
270 const MachineTraceMetrics::Ensemble &En) {
274 } // end namespace llvm