|
| 1 | +// Copyright 2016 The Go Authors. All rights reserved. |
| 2 | +// Use of this source code is governed by a BSD-style |
| 3 | +// license that can be found in the LICENSE file. |
| 4 | + |
| 5 | +package gc |
| 6 | + |
| 7 | +import ( |
| 8 | + "cmd/compile/internal/ssa" |
| 9 | + "fmt" |
| 10 | + "math" |
| 11 | +) |
| 12 | + |
| 13 | +// sparseDefState contains a Go map from ONAMEs (*Node) to sparse definition trees, and |
| 14 | +// a search helper for the CFG's dominator tree in which those definitions are embedded. |
| 15 | +// Once initialized, given a use of an ONAME within a block, the ssa definition for |
| 16 | +// that ONAME can be discovered in time roughly proportional to the log of the number |
| 17 | +// of SSA definitions of that ONAME (thus avoiding pathological quadratic behavior for |
| 18 | +// very large programs). The helper contains state (a dominator tree numbering) common |
| 19 | +// to all the sparse definition trees, as well as some necessary data obtained from |
| 20 | +// the ssa package. |
| 21 | +// |
| 22 | +// This algorithm has improved asymptotic complexity, but the constant factor is |
| 23 | +// rather large and thus it is only preferred for very large inputs containing |
| 24 | +// 1000s of blocks and variables. |
| 25 | +type sparseDefState struct { |
| 26 | + helper *ssa.SparseTreeHelper // contains one copy of information needed to do sparse mapping |
| 27 | + defmapForOname map[*Node]*onameDefs // for each ONAME, its definition set (normal and phi) |
| 28 | +} |
| 29 | + |
| 30 | +// onameDefs contains a record of definitions (ordinary and implied phi function) for a single OName. |
| 31 | +// stm is the set of definitions for the OName. |
| 32 | +// firstdef and lastuse are postorder block numberings that |
| 33 | +// conservatively bracket the entire lifetime of the OName. |
| 34 | +type onameDefs struct { |
| 35 | + stm *ssa.SparseTreeMap |
| 36 | + // firstdef and lastuse define an interval in the postorder numbering |
| 37 | + // that is guaranteed to include the entire lifetime of an ONAME. |
| 38 | + // In the postorder numbering, math.MaxInt32 is before anything, |
| 39 | + // and 0 is after-or-equal all exit nodes and infinite loops. |
| 40 | + firstdef int32 // the first definition of this ONAME *in the postorder numbering* |
| 41 | + lastuse int32 // the last use of this ONAME *in the postorder numbering* |
| 42 | +} |
| 43 | + |
| 44 | +// defsFor finds or creates-and-inserts-in-map the definition information |
| 45 | +// (sparse tree and live range) for a given OName. |
| 46 | +func (m *sparseDefState) defsFor(n *Node) *onameDefs { |
| 47 | + d := m.defmapForOname[n] |
| 48 | + if d != nil { |
| 49 | + return d |
| 50 | + } |
| 51 | + // Reminder: firstdef/lastuse are postorder indices, not block indices, |
| 52 | + // so these default values define an empty interval, not the entire one. |
| 53 | + d = &onameDefs{stm: m.helper.NewTree(), firstdef: 0, lastuse: math.MaxInt32} |
| 54 | + m.defmapForOname[n] = d |
| 55 | + return d |
| 56 | +} |
| 57 | + |
| 58 | +// Insert adds a definition at b (with specified before/within/after adjustment) |
| 59 | +// to sparse tree onameDefs. The lifetime is extended as necessary. |
| 60 | +func (m *sparseDefState) Insert(tree *onameDefs, b *ssa.Block, adjust int32) { |
| 61 | + bponum := m.helper.Ponums[b.ID] |
| 62 | + if bponum > tree.firstdef { |
| 63 | + tree.firstdef = bponum |
| 64 | + } |
| 65 | + tree.stm.Insert(b, adjust, b, m.helper) |
| 66 | +} |
| 67 | + |
| 68 | +// Use updates tree to record a use within b, extending the lifetime as necessary. |
| 69 | +func (m *sparseDefState) Use(tree *onameDefs, b *ssa.Block) { |
| 70 | + bponum := m.helper.Ponums[b.ID] |
| 71 | + if bponum < tree.lastuse { |
| 72 | + tree.lastuse = bponum |
| 73 | + } |
| 74 | +} |
| 75 | + |
| 76 | +// locatePotentialPhiFunctions finds all the places where phi functions |
| 77 | +// will be inserted into a program and records those and ordinary definitions |
| 78 | +// in a "map" (not a Go map) that given an OName and use site, returns the |
| 79 | +// SSA definition for that OName that will reach the use site (that is, |
| 80 | +// the use site's nearest def/phi site in the dominator tree.) |
| 81 | +func (s *state) locatePotentialPhiFunctions(fn *Node) *sparseDefState { |
| 82 | + // s.config.SparsePhiCutoff() is compared with product of numblocks and numvalues, |
| 83 | + // if product is smaller than cutoff, use old non-sparse method. |
| 84 | + // cutoff == 0 implies all sparse |
| 85 | + // cutoff == uint(-1) implies all non-sparse |
| 86 | + if uint64(s.f.NumValues())*uint64(s.f.NumBlocks()) < s.config.SparsePhiCutoff() { |
| 87 | + return nil |
| 88 | + } |
| 89 | + |
| 90 | + helper := ssa.NewSparseTreeHelper(s.f) |
| 91 | + po := helper.Po // index by block.ID to obtain postorder # of block. |
| 92 | + trees := make(map[*Node]*onameDefs) |
| 93 | + dm := &sparseDefState{defmapForOname: trees, helper: helper} |
| 94 | + |
| 95 | + // Process params, taking note of their special lifetimes |
| 96 | + b := s.f.Entry |
| 97 | + for _, n := range fn.Func.Dcl { |
| 98 | + switch n.Class { |
| 99 | + case PPARAM, PPARAMOUT: |
| 100 | + t := dm.defsFor(n) |
| 101 | + dm.Insert(t, b, ssa.AdjustBefore) // define param at entry block |
| 102 | + if n.Class == PPARAMOUT { |
| 103 | + dm.Use(t, po[0]) // Explicitly use PPARAMOUT at very last block |
| 104 | + } |
| 105 | + default: |
| 106 | + } |
| 107 | + } |
| 108 | + |
| 109 | + // Process memory variable. |
| 110 | + t := dm.defsFor(&memVar) |
| 111 | + dm.Insert(t, b, ssa.AdjustBefore) // define memory at entry block |
| 112 | + dm.Use(t, po[0]) // Explicitly use memory at last block |
| 113 | + |
| 114 | + // Next load the map w/ basic definitions for ONames recorded per-block |
| 115 | + // Iterate over po to avoid unreachable blocks. |
| 116 | + for i := len(po) - 1; i >= 0; i-- { |
| 117 | + b := po[i] |
| 118 | + m := s.defvars[b.ID] |
| 119 | + for n := range m { // no specified order, but per-node trees are independent. |
| 120 | + t := dm.defsFor(n) |
| 121 | + dm.Insert(t, b, ssa.AdjustWithin) |
| 122 | + } |
| 123 | + } |
| 124 | + |
| 125 | + // Find last use of each variable |
| 126 | + for _, v := range s.fwdRefs { |
| 127 | + b := v.Block |
| 128 | + name := v.Aux.(*Node) |
| 129 | + t := dm.defsFor(name) |
| 130 | + dm.Use(t, b) |
| 131 | + } |
| 132 | + |
| 133 | + for _, t := range trees { |
| 134 | + // iterating over names in the outer loop |
| 135 | + for change := true; change; { |
| 136 | + change = false |
| 137 | + for i := t.firstdef; i >= t.lastuse; i-- { |
| 138 | + // Iterating in reverse of post-order reduces number of 'change' iterations; |
| 139 | + // all possible forward flow goes through each time. |
| 140 | + b := po[i] |
| 141 | + // Within tree t, would a use at b require a phi function to ensure a single definition? |
| 142 | + // TODO: perhaps more efficient to record specific use sites instead of range? |
| 143 | + if len(b.Preds) < 2 { |
| 144 | + continue // no phi possible |
| 145 | + } |
| 146 | + phi := t.stm.Find(b, ssa.AdjustWithin, helper) // Look for defs in earlier block or AdjustBefore in this one. |
| 147 | + if phi != nil && phi.(*ssa.Block) == b { |
| 148 | + continue // has a phi already in this block. |
| 149 | + } |
| 150 | + var defseen interface{} |
| 151 | + // Do preds see different definitions? if so, need a phi function. |
| 152 | + for _, e := range b.Preds { |
| 153 | + p := e.Block() |
| 154 | + dm.Use(t, p) // always count phi pred as "use"; no-op except for loop edges, which matter. |
| 155 | + x := t.stm.Find(p, ssa.AdjustAfter, helper) // Look for defs reaching or within predecessors. |
| 156 | + if defseen == nil { |
| 157 | + defseen = x |
| 158 | + } |
| 159 | + if defseen != x || x == nil { // TODO: too conservative at loops, does better if x == nil -> continue |
| 160 | + // Need to insert a phi function here because predecessors's definitions differ. |
| 161 | + change = true |
| 162 | + // Phi insertion is at AdjustBefore, visible with find in same block at AdjustWithin or AdjustAfter. |
| 163 | + dm.Insert(t, b, ssa.AdjustBefore) |
| 164 | + break |
| 165 | + } |
| 166 | + } |
| 167 | + } |
| 168 | + } |
| 169 | + } |
| 170 | + return dm |
| 171 | +} |
| 172 | + |
| 173 | +// FindBetterDefiningBlock tries to find a better block for a definition of OName name |
| 174 | +// reaching (or within) p than p itself. If it cannot, it returns p instead. |
| 175 | +// This aids in more efficient location of phi functions, since it can skip over |
| 176 | +// branch code that might contain a definition of name if it actually does not. |
| 177 | +func (m *sparseDefState) FindBetterDefiningBlock(name *Node, p *ssa.Block) *ssa.Block { |
| 178 | + if m == nil { |
| 179 | + return p |
| 180 | + } |
| 181 | + t := m.defmapForOname[name] |
| 182 | + // For now this is fail-soft, since the old algorithm still works using the unimproved block. |
| 183 | + if t == nil { |
| 184 | + return p |
| 185 | + } |
| 186 | + x := t.stm.Find(p, ssa.AdjustAfter, m.helper) |
| 187 | + if x == nil { |
| 188 | + return p |
| 189 | + } |
| 190 | + b := x.(*ssa.Block) |
| 191 | + if b == nil { |
| 192 | + return p |
| 193 | + } |
| 194 | + return b |
| 195 | +} |
| 196 | + |
| 197 | +func (d *onameDefs) String() string { |
| 198 | + return fmt.Sprintf("onameDefs:first=%d,last=%d,tree=%s", d.firstdef, d.lastuse, d.stm.String()) |
| 199 | +} |
0 commit comments