tighten.go

Documentation: cmd/compile/internal/ssa

     1  // Copyright 2015 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 ssa
     6  
     7  import "cmd/compile/internal/base"
     8  
     9  // tighten moves Values closer to the Blocks in which they are used.
    10  // This can reduce the amount of register spilling required,
    11  // if it doesn't also create more live values.
    12  // A Value can be moved to any block that
    13  // dominates all blocks in which it is used.
    14  func tighten(f *Func) {
    15  	if base.Flag.N != 0 && len(f.Blocks) < 10000 {
    16  		// Skip the optimization in -N mode, except for huge functions.
    17  		// Too many values live across blocks can cause pathological
    18  		// behavior in the register allocator (see issue 52180).
    19  		return
    20  	}
    21  
    22  	canMove := f.Cache.allocBoolSlice(f.NumValues())
    23  	defer f.Cache.freeBoolSlice(canMove)
    24  
    25  	// Compute the memory states of each block.
    26  	startMem := f.Cache.allocValueSlice(f.NumBlocks())
    27  	defer f.Cache.freeValueSlice(startMem)
    28  	endMem := f.Cache.allocValueSlice(f.NumBlocks())
    29  	defer f.Cache.freeValueSlice(endMem)
    30  	memState(f, startMem, endMem)
    31  
    32  	for _, b := range f.Blocks {
    33  		for _, v := range b.Values {
    34  			if v.Op.isLoweredGetClosurePtr() {
    35  				// Must stay in the entry block.
    36  				continue
    37  			}
    38  			switch v.Op {
    39  			case OpPhi, OpArg, OpArgIntReg, OpArgFloatReg, OpSelect0, OpSelect1, OpSelectN:
    40  				// Phis need to stay in their block.
    41  				// Arg must stay in the entry block.
    42  				// Tuple selectors must stay with the tuple generator.
    43  				// SelectN is typically, ultimately, a register.
    44  				continue
    45  			}
    46  			// Count arguments which will need a register.
    47  			narg := 0
    48  			for _, a := range v.Args {
    49  				// SP and SB are special registers and have no effect on
    50  				// the allocation of general-purpose registers.
    51  				if a.needRegister() && a.Op != OpSB && a.Op != OpSP {
    52  					narg++
    53  				}
    54  			}
    55  			if narg >= 2 && !v.Type.IsFlags() {
    56  				// Don't move values with more than one input, as that may
    57  				// increase register pressure.
    58  				// We make an exception for flags, as we want flag generators
    59  				// moved next to uses (because we only have 1 flag register).
    60  				continue
    61  			}
    62  			canMove[v.ID] = true
    63  		}
    64  	}
    65  
    66  	// Build data structure for fast least-common-ancestor queries.
    67  	lca := makeLCArange(f)
    68  
    69  	// For each moveable value, record the block that dominates all uses found so far.
    70  	target := f.Cache.allocBlockSlice(f.NumValues())
    71  	defer f.Cache.freeBlockSlice(target)
    72  
    73  	// Grab loop information.
    74  	// We use this to make sure we don't tighten a value into a (deeper) loop.
    75  	idom := f.Idom()
    76  	loops := f.loopnest()
    77  	loops.calculateDepths()
    78  
    79  	changed := true
    80  	for changed {
    81  		changed = false
    82  
    83  		// Reset target
    84  		for i := range target {
    85  			target[i] = nil
    86  		}
    87  
    88  		// Compute target locations (for moveable values only).
    89  		// target location = the least common ancestor of all uses in the dominator tree.
    90  		for _, b := range f.Blocks {
    91  			for _, v := range b.Values {
    92  				for i, a := range v.Args {
    93  					if !canMove[a.ID] {
    94  						continue
    95  					}
    96  					use := b
    97  					if v.Op == OpPhi {
    98  						use = b.Preds[i].b
    99  					}
   100  					if target[a.ID] == nil {
   101  						target[a.ID] = use
   102  					} else {
   103  						target[a.ID] = lca.find(target[a.ID], use)
   104  					}
   105  				}
   106  			}
   107  			for _, c := range b.ControlValues() {
   108  				if !canMove[c.ID] {
   109  					continue
   110  				}
   111  				if target[c.ID] == nil {
   112  					target[c.ID] = b
   113  				} else {
   114  					target[c.ID] = lca.find(target[c.ID], b)
   115  				}
   116  			}
   117  		}
   118  
   119  		// If the target location is inside a loop,
   120  		// move the target location up to just before the loop head.
   121  		if !loops.hasIrreducible {
   122  			// Loop info might not be correct for irreducible loops. See issue 75569.
   123  			for _, b := range f.Blocks {
   124  				origloop := loops.b2l[b.ID]
   125  				for _, v := range b.Values {
   126  					t := target[v.ID]
   127  					if t == nil {
   128  						continue
   129  					}
   130  					targetloop := loops.b2l[t.ID]
   131  					for targetloop != nil && (origloop == nil || targetloop.depth > origloop.depth) {
   132  						t = idom[targetloop.header.ID]
   133  						target[v.ID] = t
   134  						targetloop = loops.b2l[t.ID]
   135  					}
   136  				}
   137  			}
   138  		}
   139  
   140  		// Move values to target locations.
   141  		for _, b := range f.Blocks {
   142  			for i := 0; i < len(b.Values); i++ {
   143  				v := b.Values[i]
   144  				t := target[v.ID]
   145  				if t == nil || t == b {
   146  					// v is not moveable, or is already in correct place.
   147  					continue
   148  				}
   149  				if mem := v.MemoryArg(); mem != nil {
   150  					if startMem[t.ID] != mem {
   151  						// We can't move a value with a memory arg unless the target block
   152  						// has that memory arg as its starting memory.
   153  						continue
   154  					}
   155  				}
   156  				if f.pass.debug > 0 {
   157  					b.Func.Warnl(v.Pos, "%v is moved", v.Op)
   158  				}
   159  				// Move v to the block which dominates its uses.
   160  				t.Values = append(t.Values, v)
   161  				v.Block = t
   162  				last := len(b.Values) - 1
   163  				b.Values[i] = b.Values[last]
   164  				b.Values[last] = nil
   165  				b.Values = b.Values[:last]
   166  				changed = true
   167  				i--
   168  			}
   169  		}
   170  	}
   171  }
   172  
   173  // phiTighten moves constants closer to phi users.
   174  // This pass avoids having lots of constants live for lots of the program.
   175  // See issue 16407.
   176  func phiTighten(f *Func) {
   177  	for _, b := range f.Blocks {
   178  		for _, v := range b.Values {
   179  			if v.Op != OpPhi {
   180  				continue
   181  			}
   182  			for i, a := range v.Args {
   183  				if !a.rematerializeable() {
   184  					continue // not a constant we can move around
   185  				}
   186  				if a.Block == b.Preds[i].b {
   187  					continue // already in the right place
   188  				}
   189  				// Make a copy of a, put in predecessor block.
   190  				v.SetArg(i, a.copyInto(b.Preds[i].b))
   191  			}
   192  		}
   193  	}
   194  }
   195  
   196  // memState computes the memory state at the beginning and end of each block of
   197  // the function. The memory state is represented by a value of mem type.
   198  // The returned result is stored in startMem and endMem, and endMem is nil for
   199  // blocks with no successors (Exit,Ret,RetJmp blocks). This algorithm is not
   200  // suitable for infinite loop blocks that do not contain any mem operations.
   201  // For example:
   202  // b1:
   203  //
   204  //	(some values)
   205  //
   206  // plain -> b2
   207  // b2: <- b1 b2
   208  // Plain -> b2
   209  //
   210  // Algorithm introduction:
   211  //  1. The start memory state of a block is InitMem, a Phi node of type mem or
   212  //     an incoming memory value.
   213  //  2. The start memory state of a block is consistent with the end memory state
   214  //     of its parent nodes. If the start memory state of a block is a Phi value,
   215  //     then the end memory state of its parent nodes is consistent with the
   216  //     corresponding argument value of the Phi node.
   217  //  3. The algorithm first obtains the memory state of some blocks in the tree
   218  //     in the first step. Then floods the known memory state to other nodes in
   219  //     the second step.
   220  func memState(f *Func, startMem, endMem []*Value) {
   221  	// This slice contains the set of blocks that have had their startMem set but this
   222  	// startMem value has not yet been propagated to the endMem of its predecessors
   223  	changed := make([]*Block, 0)
   224  	// First step, init the memory state of some blocks.
   225  	for _, b := range f.Blocks {
   226  		for _, v := range b.Values {
   227  			var mem *Value
   228  			if v.Op == OpPhi {
   229  				if v.Type.IsMemory() {
   230  					mem = v
   231  				}
   232  			} else if v.Op == OpInitMem {
   233  				mem = v // This is actually not needed.
   234  			} else if a := v.MemoryArg(); a != nil && a.Block != b {
   235  				// The only incoming memory value doesn't belong to this block.
   236  				mem = a
   237  			}
   238  			if mem != nil {
   239  				if old := startMem[b.ID]; old != nil {
   240  					if old == mem {
   241  						continue
   242  					}
   243  					f.Fatalf("func %s, startMem[%v] has different values, old %v, new %v", f.Name, b, old, mem)
   244  				}
   245  				startMem[b.ID] = mem
   246  				changed = append(changed, b)
   247  			}
   248  		}
   249  	}
   250  
   251  	// Second step, floods the known memory state of some blocks to others.
   252  	for len(changed) != 0 {
   253  		top := changed[0]
   254  		changed = changed[1:]
   255  		mem := startMem[top.ID]
   256  		for i, p := range top.Preds {
   257  			pb := p.b
   258  			if endMem[pb.ID] != nil {
   259  				continue
   260  			}
   261  			if mem.Op == OpPhi && mem.Block == top {
   262  				endMem[pb.ID] = mem.Args[i]
   263  			} else {
   264  				endMem[pb.ID] = mem
   265  			}
   266  			if startMem[pb.ID] == nil {
   267  				startMem[pb.ID] = endMem[pb.ID]
   268  				changed = append(changed, pb)
   269  			}
   270  		}
   271  	}
   272  }
   273
View as plain text