fuse.go

     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 (
     8  	"cmd/internal/src"
     9  	"fmt"
    10  )
    11  
    12  // fuseEarly runs fuse(f, fuseTypePlain|fuseTypeIntInRange).
    13  func fuseEarly(f *Func) { fuse(f, fuseTypePlain|fuseTypeIntInRange) }
    14  
    15  // fuseLate runs fuse(f, fuseTypePlain|fuseTypeIf|fuseTypeBranchRedirect).
    16  func fuseLate(f *Func) { fuse(f, fuseTypePlain|fuseTypeIf|fuseTypeBranchRedirect) }
    17  
    18  type fuseType uint8
    19  
    20  const (
    21  	fuseTypePlain fuseType = 1 << iota
    22  	fuseTypeIf
    23  	fuseTypeIntInRange
    24  	fuseTypeBranchRedirect
    25  	fuseTypeShortCircuit
    26  )
    27  
    28  // fuse simplifies control flow by joining basic blocks.
    29  func fuse(f *Func, typ fuseType) {
    30  	for changed := true; changed; {
    31  		changed = false
    32  		// Be sure to avoid quadratic behavior in fuseBlockPlain. See issue 13554.
    33  		// Previously this was dealt with using backwards iteration, now fuseBlockPlain
    34  		// handles large runs of blocks.
    35  		for i := len(f.Blocks) - 1; i >= 0; i-- {
    36  			b := f.Blocks[i]
    37  			if typ&fuseTypeIf != 0 {
    38  				changed = fuseBlockIf(b) || changed
    39  			}
    40  			if typ&fuseTypeIntInRange != 0 {
    41  				changed = fuseIntegerComparisons(b) || changed
    42  			}
    43  			if typ&fuseTypePlain != 0 {
    44  				changed = fuseBlockPlain(b) || changed
    45  			}
    46  			if typ&fuseTypeShortCircuit != 0 {
    47  				changed = shortcircuitBlock(b) || changed
    48  			}
    49  		}
    50  
    51  		if typ&fuseTypeBranchRedirect != 0 {
    52  			changed = fuseBranchRedirect(f) || changed
    53  		}
    54  		if changed {
    55  			f.invalidateCFG()
    56  		}
    57  	}
    58  }
    59  
    60  // fuseBlockIf handles the following cases where s0 and s1 are empty blocks.
    61  //
    62  //	   b        b           b       b
    63  //	\ / \ /    | \  /    \ / |     | |
    64  //	 s0  s1    |  s1      s0 |     | |
    65  //	  \ /      | /         \ |     | |
    66  //	   ss      ss           ss      ss
    67  //
    68  // If all Phi ops in ss have identical variables for slots corresponding to
    69  // s0, s1 and b then the branch can be dropped.
    70  // This optimization often comes up in switch statements with multiple
    71  // expressions in a case clause:
    72  //
    73  //	switch n {
    74  //	  case 1,2,3: return 4
    75  //	}
    76  //
    77  // TODO: If ss doesn't contain any OpPhis, are s0 and s1 dead code anyway.
    78  func fuseBlockIf(b *Block) bool {
    79  	if b.Kind != BlockIf {
    80  		return false
    81  	}
    82  	// It doesn't matter how much Preds does s0 or s1 have.
    83  	var ss0, ss1 *Block
    84  	s0 := b.Succs[0].b
    85  	i0 := b.Succs[0].i
    86  	if s0.Kind != BlockPlain || !isEmpty(s0) {
    87  		s0, ss0 = b, s0
    88  	} else {
    89  		ss0 = s0.Succs[0].b
    90  		i0 = s0.Succs[0].i
    91  	}
    92  	s1 := b.Succs[1].b
    93  	i1 := b.Succs[1].i
    94  	if s1.Kind != BlockPlain || !isEmpty(s1) {
    95  		s1, ss1 = b, s1
    96  	} else {
    97  		ss1 = s1.Succs[0].b
    98  		i1 = s1.Succs[0].i
    99  	}
   100  	if ss0 != ss1 {
   101  		if s0.Kind == BlockPlain && isEmpty(s0) && s1.Kind == BlockPlain && isEmpty(s1) {
   102  			// Two special cases where both s0, s1 and ss are empty blocks.
   103  			if s0 == ss1 {
   104  				s0, ss0 = b, ss1
   105  			} else if ss0 == s1 {
   106  				s1, ss1 = b, ss0
   107  			} else {
   108  				return false
   109  			}
   110  		} else {
   111  			return false
   112  		}
   113  	}
   114  	ss := ss0
   115  
   116  	// s0 and s1 are equal with b if the corresponding block is missing
   117  	// (2nd, 3rd and 4th case in the figure).
   118  
   119  	for _, v := range ss.Values {
   120  		if v.Op == OpPhi && v.Uses > 0 && v.Args[i0] != v.Args[i1] {
   121  			return false
   122  		}
   123  	}
   124  
   125  	// We do not need to redirect the Preds of s0 and s1 to ss,
   126  	// the following optimization will do this.
   127  	b.removeEdge(0)
   128  	if s0 != b && len(s0.Preds) == 0 {
   129  		s0.removeEdge(0)
   130  		// Move any (dead) values in s0 to b,
   131  		// where they will be eliminated by the next deadcode pass.
   132  		for _, v := range s0.Values {
   133  			v.Block = b
   134  		}
   135  		b.Values = append(b.Values, s0.Values...)
   136  		// Clear s0.
   137  		s0.Kind = BlockInvalid
   138  		s0.Values = nil
   139  		s0.Succs = nil
   140  		s0.Preds = nil
   141  	}
   142  
   143  	b.Kind = BlockPlain
   144  	b.Likely = BranchUnknown
   145  	b.ResetControls()
   146  	// The values in b may be dead codes, and clearing them in time may
   147  	// obtain new optimization opportunities.
   148  	// First put dead values that can be deleted into a slice walkValues.
   149  	// Then put their arguments in walkValues before resetting the dead values
   150  	// in walkValues, because the arguments may also become dead values.
   151  	walkValues := []*Value{}
   152  	for _, v := range b.Values {
   153  		if v.Uses == 0 && v.removeable() {
   154  			walkValues = append(walkValues, v)
   155  		}
   156  	}
   157  	for len(walkValues) != 0 {
   158  		v := walkValues[len(walkValues)-1]
   159  		walkValues = walkValues[:len(walkValues)-1]
   160  		if v.Uses == 0 && v.removeable() {
   161  			walkValues = append(walkValues, v.Args...)
   162  			v.reset(OpInvalid)
   163  		}
   164  	}
   165  	return true
   166  }
   167  
   168  // isEmpty reports whether b contains any live values.
   169  // There may be false positives.
   170  func isEmpty(b *Block) bool {
   171  	for _, v := range b.Values {
   172  		if v.Uses > 0 || v.Op.IsCall() || v.Op.HasSideEffects() || v.Type.IsVoid() || opcodeTable[v.Op].nilCheck {
   173  			return false
   174  		}
   175  	}
   176  	return true
   177  }
   178  
   179  // fuseBlockPlain handles a run of blocks with length >= 2,
   180  // whose interior has single predecessors and successors,
   181  // b must be BlockPlain, allowing it to be any node except the
   182  // last (multiple successors means not BlockPlain).
   183  // Cycles are handled and merged into b's successor.
   184  func fuseBlockPlain(b *Block) bool {
   185  	if b.Kind != BlockPlain {
   186  		return false
   187  	}
   188  
   189  	c := b.Succs[0].b
   190  	if len(c.Preds) != 1 || c == b { // At least 2 distinct blocks.
   191  		return false
   192  	}
   193  
   194  	// find earliest block in run.  Avoid simple cycles.
   195  	for len(b.Preds) == 1 && b.Preds[0].b != c && b.Preds[0].b.Kind == BlockPlain {
   196  		b = b.Preds[0].b
   197  	}
   198  
   199  	// find latest block in run.  Still beware of simple cycles.
   200  	for {
   201  		if c.Kind != BlockPlain {
   202  			break
   203  		} // Has exactly 1 successor
   204  		cNext := c.Succs[0].b
   205  		if cNext == b {
   206  			break
   207  		} // not a cycle
   208  		if len(cNext.Preds) != 1 {
   209  			break
   210  		} // no other incoming edge
   211  		c = cNext
   212  	}
   213  
   214  	// Try to preserve any statement marks on the ends of blocks; move values to C
   215  	var b_next *Block
   216  	for bx := b; bx != c; bx = b_next {
   217  		// For each bx with an end-of-block statement marker,
   218  		// try to move it to a value in the next block,
   219  		// or to the next block's end, if possible.
   220  		b_next = bx.Succs[0].b
   221  		if bx.Pos.IsStmt() == src.PosIsStmt {
   222  			l := bx.Pos.Line() // looking for another place to mark for line l
   223  			outOfOrder := false
   224  			for _, v := range b_next.Values {
   225  				if v.Pos.IsStmt() == src.PosNotStmt {
   226  					continue
   227  				}
   228  				if l == v.Pos.Line() { // Found a Value with same line, therefore done.
   229  					v.Pos = v.Pos.WithIsStmt()
   230  					l = 0
   231  					break
   232  				}
   233  				if l < v.Pos.Line() {
   234  					// The order of values in a block is not specified so OOO in a block is not interesting,
   235  					// but they do all come before the end of the block, so this disqualifies attaching to end of b_next.
   236  					outOfOrder = true
   237  				}
   238  			}
   239  			if l != 0 && !outOfOrder && (b_next.Pos.Line() == l || b_next.Pos.IsStmt() != src.PosIsStmt) {
   240  				b_next.Pos = bx.Pos.WithIsStmt()
   241  			}
   242  		}
   243  		// move all of bx's values to c (note containing loop excludes c)
   244  		for _, v := range bx.Values {
   245  			v.Block = c
   246  		}
   247  	}
   248  
   249  	// Compute the total number of values and find the largest value slice in the run, to maximize chance of storage reuse.
   250  	total := 0
   251  	totalBeforeMax := 0 // number of elements preceding the maximum block (i.e. its position in the result).
   252  	max_b := b          // block with maximum capacity
   253  
   254  	for bx := b; ; bx = bx.Succs[0].b {
   255  		if cap(bx.Values) > cap(max_b.Values) {
   256  			totalBeforeMax = total
   257  			max_b = bx
   258  		}
   259  		total += len(bx.Values)
   260  		if bx == c {
   261  			break
   262  		}
   263  	}
   264  
   265  	// Use c's storage if fused blocks will fit, else use the max if that will fit, else allocate new storage.
   266  
   267  	// Take care to avoid c.Values pointing to b.valstorage.
   268  	// See golang.org/issue/18602.
   269  
   270  	// It's important to keep the elements in the same order; maintenance of
   271  	// debugging information depends on the order of *Values in Blocks.
   272  	// This can also cause changes in the order (which may affect other
   273  	// optimizations and possibly compiler output) for 32-vs-64 bit compilation
   274  	// platforms (word size affects allocation bucket size affects slice capacity).
   275  
   276  	// figure out what slice will hold the values,
   277  	// preposition the destination elements if not allocating new storage
   278  	var t []*Value
   279  	if total <= len(c.valstorage) {
   280  		t = c.valstorage[:total]
   281  		max_b = c
   282  		totalBeforeMax = total - len(c.Values)
   283  		copy(t[totalBeforeMax:], c.Values)
   284  	} else if total <= cap(max_b.Values) { // in place, somewhere
   285  		t = max_b.Values[0:total]
   286  		copy(t[totalBeforeMax:], max_b.Values)
   287  	} else {
   288  		t = make([]*Value, total)
   289  		max_b = nil
   290  	}
   291  
   292  	// copy the values
   293  	copyTo := 0
   294  	for bx := b; ; bx = bx.Succs[0].b {
   295  		if bx != max_b {
   296  			copy(t[copyTo:], bx.Values)
   297  		} else if copyTo != totalBeforeMax { // trust but verify.
   298  			panic(fmt.Errorf("totalBeforeMax (%d) != copyTo (%d), max_b=%v, b=%v, c=%v", totalBeforeMax, copyTo, max_b, b, c))
   299  		}
   300  		if bx == c {
   301  			break
   302  		}
   303  		copyTo += len(bx.Values)
   304  	}
   305  	c.Values = t
   306  
   307  	// replace b->c edge with preds(b) -> c
   308  	c.predstorage[0] = Edge{}
   309  	if len(b.Preds) > len(b.predstorage) {
   310  		c.Preds = b.Preds
   311  	} else {
   312  		c.Preds = append(c.predstorage[:0], b.Preds...)
   313  	}
   314  	for i, e := range c.Preds {
   315  		p := e.b
   316  		p.Succs[e.i] = Edge{c, i}
   317  	}
   318  	f := b.Func
   319  	if f.Entry == b {
   320  		f.Entry = c
   321  	}
   322  
   323  	// trash b's fields, just in case
   324  	for bx := b; bx != c; bx = b_next {
   325  		b_next = bx.Succs[0].b
   326  
   327  		bx.Kind = BlockInvalid
   328  		bx.Values = nil
   329  		bx.Preds = nil
   330  		bx.Succs = nil
   331  	}
   332  	return true
   333  }
   334
View as plain text