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 /// Valid when hasValidDepth().
109 const MachineBasicBlock *Pred;
111 /// Trace successor, or NULL for the last block in the trace.
112 /// Valid when hasValidHeight().
113 const MachineBasicBlock *Succ;
115 /// The block number of the head of the trace. (When hasValidDepth()).
118 /// The block number of the tail of the trace. (When hasValidHeight()).
121 /// Accumulated number of instructions in the trace above this block.
122 /// Does not include instructions in this block.
125 /// Accumulated number of instructions in the trace below this block.
126 /// Includes instructions in this block.
127 unsigned InstrHeight;
129 TraceBlockInfo() : Pred(0), Succ(0), InstrDepth(~0u), InstrHeight(~0u) {}
131 /// Returns true if the depth resources have been computed from the trace
132 /// above this block.
133 bool hasValidDepth() const { return InstrDepth != ~0u; }
135 /// Returns true if the height resources have been computed from the trace
136 /// below this block.
137 bool hasValidHeight() const { return InstrHeight != ~0u; }
139 /// Invalidate depth resources when some block above this one has changed.
140 void invalidateDepth() { InstrDepth = ~0u; }
142 /// Invalidate height resources when a block below this one has changed.
143 void invalidateHeight() { InstrHeight = ~0u; }
145 void print(raw_ostream&) const;
148 /// A trace represents a plausible sequence of executed basic blocks that
149 /// passes through the current basic block one. The Trace class serves as a
150 /// handle to internal cached data structures.
156 explicit Trace(Ensemble &te, TraceBlockInfo &tbi) : TE(te), TBI(tbi) {}
157 void print(raw_ostream&) const;
159 /// Compute the total number of instructions in the trace.
160 unsigned getInstrCount() const {
161 return TBI.InstrDepth + TBI.InstrHeight;
165 /// A trace ensemble is a collection of traces selected using the same
166 /// strategy, for example 'minimum resource height'. There is one trace for
167 /// every block in the function.
169 SmallVector<TraceBlockInfo, 4> BlockInfo;
172 void computeTrace(const MachineBasicBlock*);
173 void computeDepthResources(const MachineBasicBlock*);
174 void computeHeightResources(const MachineBasicBlock*);
177 MachineTraceMetrics &CT;
178 virtual const MachineBasicBlock *pickTracePred(const MachineBasicBlock*) =0;
179 virtual const MachineBasicBlock *pickTraceSucc(const MachineBasicBlock*) =0;
180 explicit Ensemble(MachineTraceMetrics*);
181 MachineLoop *getLoopFor(const MachineBasicBlock*);
182 const TraceBlockInfo *getDepthResources(const MachineBasicBlock*) const;
183 const TraceBlockInfo *getHeightResources(const MachineBasicBlock*) const;
187 virtual const char *getName() const =0;
188 void print(raw_ostream&) const;
189 void invalidate(const MachineBasicBlock *MBB);
191 /// Get the trace that passes through MBB.
192 /// The trace is computed on demand.
193 Trace getTrace(const MachineBasicBlock *MBB);
196 /// Strategies for selecting traces.
198 /// Select the trace through a block that has the fewest instructions.
204 /// Get the trace ensemble representing the given trace selection strategy.
205 /// The returned Ensemble object is owned by the MachineTraceMetrics analysis,
206 /// and valid for the lifetime of the analysis pass.
207 Ensemble *getEnsemble(Strategy);
209 /// Invalidate cached information about MBB. This must be called *before* MBB
210 /// is erased, or the CFG is otherwise changed.
211 void invalidate(const MachineBasicBlock *MBB);
214 // One entry per basic block, indexed by block number.
215 SmallVector<FixedBlockInfo, 4> BlockInfo;
217 // One ensemble per strategy.
218 Ensemble* Ensembles[TS_NumStrategies];
221 inline raw_ostream &operator<<(raw_ostream &OS,
222 const MachineTraceMetrics::Trace &Tr) {
227 inline raw_ostream &operator<<(raw_ostream &OS,
228 const MachineTraceMetrics::Ensemble &En) {
232 } // end namespace llvm