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/SmallVector.h"
51 #include "llvm/CodeGen/MachineFunctionPass.h"
55 class TargetInstrInfo;
56 class TargetRegisterInfo;
57 class MachineBasicBlock;
58 class MachineRegisterInfo;
59 class MachineLoopInfo;
63 class MachineTraceMetrics : public MachineFunctionPass {
64 const TargetInstrInfo *TII;
65 const TargetRegisterInfo *TRI;
66 const MachineRegisterInfo *MRI;
67 const MachineLoopInfo *Loops;
73 MachineTraceMetrics();
74 void getAnalysisUsage(AnalysisUsage&) const;
75 bool runOnMachineFunction(MachineFunction&);
78 friend class Ensemble;
81 /// Per-basic block information that doesn't depend on the trace through the
83 struct FixedBlockInfo {
84 /// The number of non-trivial instructions in the block.
85 /// Doesn't count PHI and COPY instructions that are likely to be removed.
88 /// True when the block contains calls.
91 FixedBlockInfo() : InstrCount(~0u), HasCalls(false) {}
93 /// Returns true when resource information for this block has been computed.
94 bool hasResources() const { return InstrCount != ~0u; }
96 /// Invalidate resource information.
97 void invalidate() { InstrCount = ~0u; }
100 /// Get the fixed resource information about MBB. Compute it on demand.
101 const FixedBlockInfo *getResources(const MachineBasicBlock*);
103 /// Per-basic block information that relates to a specific trace through the
104 /// block. Convergent traces means that only one of these is required per
105 /// block in a trace ensemble.
106 struct TraceBlockInfo {
107 /// Trace predecessor, or NULL for the first block in the trace.
108 const MachineBasicBlock *Pred;
110 /// Trace successor, or NULL for the last block in the trace.
111 const MachineBasicBlock *Succ;
113 /// The block number of the head of the trace. (When hasValidDepth()).
116 /// The block number of the tail of the trace. (When hasValidHeight()).
119 /// Accumulated number of instructions in the trace above this block.
120 /// Does not include instructions in this block.
123 /// Accumulated number of instructions in the trace below this block.
124 /// Includes instructions in this block.
125 unsigned InstrHeight;
127 TraceBlockInfo() : Pred(0), Succ(0), InstrDepth(~0u), InstrHeight(~0u) {}
129 /// Returns true if the depth resources have been computed from the trace
130 /// above this block.
131 bool hasValidDepth() const { return InstrDepth != ~0u; }
133 /// Returns true if the height resources have been computed from the trace
134 /// below this block.
135 bool hasValidHeight() const { return InstrHeight != ~0u; }
137 /// Invalidate depth resources when some block above this one has changed.
138 void invalidateDepth() { InstrDepth = ~0u; }
140 /// Invalidate height resources when a block below this one has changed.
141 void invalidateHeight() { InstrHeight = ~0u; }
144 /// A trace represents a plausible sequence of executed basic blocks that
145 /// passes through the current basic block one. The Trace class serves as a
146 /// handle to internal cached data structures.
152 explicit Trace(Ensemble &te, TraceBlockInfo &tbi) : TE(te), TBI(tbi) {}
153 void print(raw_ostream&) const;
155 /// Compute the total number of instructions in the trace.
156 unsigned getInstrCount() const {
157 return TBI.InstrDepth + TBI.InstrHeight;
161 /// A trace ensemble is a collection of traces selected using the same
162 /// strategy, for example 'minimum resource height'. There is one trace for
163 /// every block in the function.
165 SmallVector<TraceBlockInfo, 4> BlockInfo;
168 void computeTrace(const MachineBasicBlock*);
169 void computeDepthResources(const MachineBasicBlock*);
170 void computeHeightResources(const MachineBasicBlock*);
173 MachineTraceMetrics &CT;
174 virtual const MachineBasicBlock *pickTracePred(const MachineBasicBlock*) =0;
175 virtual const MachineBasicBlock *pickTraceSucc(const MachineBasicBlock*) =0;
176 explicit Ensemble(MachineTraceMetrics*);
177 MachineLoop *getLoopFor(const MachineBasicBlock*);
178 const TraceBlockInfo *getDepthResources(const MachineBasicBlock*) const;
179 const TraceBlockInfo *getHeightResources(const MachineBasicBlock*) const;
183 virtual const char *getName() =0;
184 void invalidate(const MachineBasicBlock *MBB);
186 /// Get the trace that passes through MBB.
187 /// The trace is computed on demand.
188 Trace getTrace(const MachineBasicBlock *MBB);
191 /// Strategies for selecting traces.
193 /// Select the trace through a block that has the fewest instructions.
199 /// Get the trace ensemble representing the given trace selection strategy.
200 /// The returned Ensemble object is owned by the MachineTraceMetrics analysis,
201 /// and valid for the lifetime of the analysis pass.
202 Ensemble *getEnsemble(Strategy);
204 /// Invalidate cached information about MBB. This must be called *before* MBB
205 /// is erased, or the CFG is otherwise changed.
206 void invalidate(const MachineBasicBlock *MBB);
209 // One entry per basic block, indexed by block number.
210 SmallVector<FixedBlockInfo, 4> BlockInfo;
212 // One ensemble per strategy.
213 Ensemble* Ensembles[TS_NumStrategies];
216 inline raw_ostream &operator<<(raw_ostream &OS,
217 const MachineTraceMetrics::Trace &Tr) {
222 } // end namespace llvm