schedule.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/compile/internal/base"
     9  	"cmd/compile/internal/types"
    10  	"cmp"
    11  	"container/heap"
    12  	"slices"
    13  	"sort"
    14  )
    15  
    16  const (
    17  	ScorePhi       = iota // towards top of block
    18  	ScoreArg              // must occur at the top of the entry block
    19  	ScoreInitMem          // after the args - used as mark by debug info generation
    20  	ScoreReadTuple        // must occur immediately after tuple-generating insn (or call)
    21  	ScoreNilCheck
    22  	ScoreMemory
    23  	ScoreReadFlags
    24  	ScoreDefault
    25  	ScoreFlags
    26  	ScoreControl // towards bottom of block
    27  )
    28  
    29  type ValHeap struct {
    30  	a           []*Value
    31  	score       []int8
    32  	inBlockUses []bool
    33  }
    34  
    35  func (h ValHeap) Len() int      { return len(h.a) }
    36  func (h ValHeap) Swap(i, j int) { a := h.a; a[i], a[j] = a[j], a[i] }
    37  
    38  func (h *ValHeap) Push(x interface{}) {
    39  	// Push and Pop use pointer receivers because they modify the slice's length,
    40  	// not just its contents.
    41  	v := x.(*Value)
    42  	h.a = append(h.a, v)
    43  }
    44  func (h *ValHeap) Pop() interface{} {
    45  	old := h.a
    46  	n := len(old)
    47  	x := old[n-1]
    48  	h.a = old[0 : n-1]
    49  	return x
    50  }
    51  func (h ValHeap) Less(i, j int) bool {
    52  	x := h.a[i]
    53  	y := h.a[j]
    54  	sx := h.score[x.ID]
    55  	sy := h.score[y.ID]
    56  	if c := sx - sy; c != 0 {
    57  		return c < 0 // lower scores come earlier.
    58  	}
    59  	// Note: only scores are required for correct scheduling.
    60  	// Everything else is just heuristics.
    61  
    62  	ix := h.inBlockUses[x.ID]
    63  	iy := h.inBlockUses[y.ID]
    64  	if ix != iy {
    65  		return ix // values with in-block uses come earlier
    66  	}
    67  
    68  	if x.Pos != y.Pos { // Favor in-order line stepping
    69  		return x.Pos.Before(y.Pos)
    70  	}
    71  	if x.Op != OpPhi {
    72  		if c := len(x.Args) - len(y.Args); c != 0 {
    73  			return c > 0 // smaller args come later
    74  		}
    75  	}
    76  	if c := x.Uses - y.Uses; c != 0 {
    77  		return c > 0 // smaller uses come later
    78  	}
    79  	// These comparisons are fairly arbitrary.
    80  	// The goal here is stability in the face
    81  	// of unrelated changes elsewhere in the compiler.
    82  	if c := x.AuxInt - y.AuxInt; c != 0 {
    83  		return c < 0
    84  	}
    85  	if cmp := x.Type.Compare(y.Type); cmp != types.CMPeq {
    86  		return cmp == types.CMPlt
    87  	}
    88  	return x.ID < y.ID
    89  }
    90  
    91  func (op Op) isLoweredGetClosurePtr() bool {
    92  	switch op {
    93  	case OpAMD64LoweredGetClosurePtr, OpPPC64LoweredGetClosurePtr, OpARMLoweredGetClosurePtr, OpARM64LoweredGetClosurePtr,
    94  		Op386LoweredGetClosurePtr, OpMIPS64LoweredGetClosurePtr, OpLOONG64LoweredGetClosurePtr, OpS390XLoweredGetClosurePtr, OpMIPSLoweredGetClosurePtr,
    95  		OpRISCV64LoweredGetClosurePtr, OpWasmLoweredGetClosurePtr:
    96  		return true
    97  	}
    98  	return false
    99  }
   100  
   101  // Schedule the Values in each Block. After this phase returns, the
   102  // order of b.Values matters and is the order in which those values
   103  // will appear in the assembly output. For now it generates a
   104  // reasonable valid schedule using a priority queue. TODO(khr):
   105  // schedule smarter.
   106  func schedule(f *Func) {
   107  	// reusable priority queue
   108  	priq := new(ValHeap)
   109  
   110  	// "priority" for a value
   111  	score := f.Cache.allocInt8Slice(f.NumValues())
   112  	defer f.Cache.freeInt8Slice(score)
   113  
   114  	// maps mem values to the next live memory value
   115  	nextMem := f.Cache.allocValueSlice(f.NumValues())
   116  	defer f.Cache.freeValueSlice(nextMem)
   117  
   118  	// inBlockUses records whether a value is used in the block
   119  	// in which it lives. (block control values don't count as uses.)
   120  	inBlockUses := f.Cache.allocBoolSlice(f.NumValues())
   121  	defer f.Cache.freeBoolSlice(inBlockUses)
   122  	if f.Config.optimize {
   123  		for _, b := range f.Blocks {
   124  			for _, v := range b.Values {
   125  				for _, a := range v.Args {
   126  					if a.Block == b {
   127  						inBlockUses[a.ID] = true
   128  					}
   129  				}
   130  			}
   131  		}
   132  	}
   133  	priq.inBlockUses = inBlockUses
   134  
   135  	for _, b := range f.Blocks {
   136  		// Compute score. Larger numbers are scheduled closer to the end of the block.
   137  		for _, v := range b.Values {
   138  			switch {
   139  			case v.Op.isLoweredGetClosurePtr():
   140  				// We also score GetLoweredClosurePtr as early as possible to ensure that the
   141  				// context register is not stomped. GetLoweredClosurePtr should only appear
   142  				// in the entry block where there are no phi functions, so there is no
   143  				// conflict or ambiguity here.
   144  				if b != f.Entry {
   145  					f.Fatalf("LoweredGetClosurePtr appeared outside of entry block, b=%s", b.String())
   146  				}
   147  				score[v.ID] = ScorePhi
   148  			case opcodeTable[v.Op].nilCheck:
   149  				// Nil checks must come before loads from the same address.
   150  				score[v.ID] = ScoreNilCheck
   151  			case v.Op == OpPhi:
   152  				// We want all the phis first.
   153  				score[v.ID] = ScorePhi
   154  			case v.Op == OpArgIntReg || v.Op == OpArgFloatReg:
   155  				// In-register args must be scheduled as early as possible to ensure that they
   156  				// are not stomped (similar to the closure pointer above).
   157  				// In particular, they need to come before regular OpArg operations because
   158  				// of how regalloc places spill code (see regalloc.go:placeSpills:mustBeFirst).
   159  				if b != f.Entry {
   160  					f.Fatalf("%s appeared outside of entry block, b=%s", v.Op, b.String())
   161  				}
   162  				score[v.ID] = ScorePhi
   163  			case v.Op == OpArg || v.Op == OpSP || v.Op == OpSB:
   164  				// We want all the args as early as possible, for better debugging.
   165  				score[v.ID] = ScoreArg
   166  			case v.Op == OpInitMem:
   167  				// Early, but after args. See debug.go:buildLocationLists
   168  				score[v.ID] = ScoreInitMem
   169  			case v.Type.IsMemory():
   170  				// Schedule stores as early as possible. This tends to
   171  				// reduce register pressure.
   172  				score[v.ID] = ScoreMemory
   173  			case v.Op == OpSelect0 || v.Op == OpSelect1 || v.Op == OpSelectN:
   174  				// Tuple selectors need to appear immediately after the instruction
   175  				// that generates the tuple.
   176  				score[v.ID] = ScoreReadTuple
   177  			case v.hasFlagInput():
   178  				// Schedule flag-reading ops earlier, to minimize the lifetime
   179  				// of flag values.
   180  				score[v.ID] = ScoreReadFlags
   181  			case v.isFlagOp():
   182  				// Schedule flag register generation as late as possible.
   183  				// This makes sure that we only have one live flags
   184  				// value at a time.
   185  				// Note that this case is after the case above, so values
   186  				// which both read and generate flags are given ScoreReadFlags.
   187  				score[v.ID] = ScoreFlags
   188  			default:
   189  				score[v.ID] = ScoreDefault
   190  				// If we're reading flags, schedule earlier to keep flag lifetime short.
   191  				for _, a := range v.Args {
   192  					if a.isFlagOp() {
   193  						score[v.ID] = ScoreReadFlags
   194  					}
   195  				}
   196  			}
   197  		}
   198  		for _, c := range b.ControlValues() {
   199  			// Force the control values to be scheduled at the end,
   200  			// unless they have other special priority.
   201  			if c.Block != b || score[c.ID] < ScoreReadTuple {
   202  				continue
   203  			}
   204  			if score[c.ID] == ScoreReadTuple {
   205  				score[c.Args[0].ID] = ScoreControl
   206  				continue
   207  			}
   208  			score[c.ID] = ScoreControl
   209  		}
   210  	}
   211  	priq.score = score
   212  
   213  	// An edge represents a scheduling constraint that x must appear before y in the schedule.
   214  	type edge struct {
   215  		x, y *Value
   216  	}
   217  	edges := make([]edge, 0, 64)
   218  
   219  	// inEdges is the number of scheduling edges incoming from values that haven't been scheduled yet.
   220  	// i.e. inEdges[y.ID] = |e in edges where e.y == y and e.x is not in the schedule yet|.
   221  	inEdges := f.Cache.allocInt32Slice(f.NumValues())
   222  	defer f.Cache.freeInt32Slice(inEdges)
   223  
   224  	for _, b := range f.Blocks {
   225  		edges = edges[:0]
   226  		// Standard edges: from the argument of a value to that value.
   227  		for _, v := range b.Values {
   228  			if v.Op == OpPhi {
   229  				// If a value is used by a phi, it does not induce
   230  				// a scheduling edge because that use is from the
   231  				// previous iteration.
   232  				continue
   233  			}
   234  			for _, a := range v.Args {
   235  				if a.Block == b {
   236  					edges = append(edges, edge{a, v})
   237  				}
   238  			}
   239  		}
   240  
   241  		// Find store chain for block.
   242  		// Store chains for different blocks overwrite each other, so
   243  		// the calculated store chain is good only for this block.
   244  		for _, v := range b.Values {
   245  			if v.Op != OpPhi && v.Op != OpInitMem && v.Type.IsMemory() {
   246  				nextMem[v.MemoryArg().ID] = v
   247  			}
   248  		}
   249  
   250  		// Add edges to enforce that any load must come before the following store.
   251  		for _, v := range b.Values {
   252  			if v.Op == OpPhi || v.Type.IsMemory() {
   253  				continue
   254  			}
   255  			w := v.MemoryArg()
   256  			if w == nil {
   257  				continue
   258  			}
   259  			if s := nextMem[w.ID]; s != nil && s.Block == b {
   260  				edges = append(edges, edge{v, s})
   261  			}
   262  		}
   263  
   264  		// Sort all the edges by source Value ID.
   265  		slices.SortFunc(edges, func(a, b edge) int {
   266  			return cmp.Compare(a.x.ID, b.x.ID)
   267  		})
   268  		// Compute inEdges for values in this block.
   269  		for _, e := range edges {
   270  			inEdges[e.y.ID]++
   271  		}
   272  
   273  		// Initialize priority queue with schedulable values.
   274  		priq.a = priq.a[:0]
   275  		for _, v := range b.Values {
   276  			if inEdges[v.ID] == 0 {
   277  				heap.Push(priq, v)
   278  			}
   279  		}
   280  
   281  		// Produce the schedule. Pick the highest priority scheduleable value,
   282  		// add it to the schedule, add any of its uses that are now scheduleable
   283  		// to the queue, and repeat.
   284  		nv := len(b.Values)
   285  		b.Values = b.Values[:0]
   286  		for priq.Len() > 0 {
   287  			// Schedule the next schedulable value in priority order.
   288  			v := heap.Pop(priq).(*Value)
   289  			b.Values = append(b.Values, v)
   290  
   291  			// Find all the scheduling edges out from this value.
   292  			i := sort.Search(len(edges), func(i int) bool {
   293  				return edges[i].x.ID >= v.ID
   294  			})
   295  			j := sort.Search(len(edges), func(i int) bool {
   296  				return edges[i].x.ID > v.ID
   297  			})
   298  			// Decrement inEdges for each target of edges from v.
   299  			for _, e := range edges[i:j] {
   300  				inEdges[e.y.ID]--
   301  				if inEdges[e.y.ID] == 0 {
   302  					heap.Push(priq, e.y)
   303  				}
   304  			}
   305  		}
   306  		if len(b.Values) != nv {
   307  			f.Fatalf("schedule does not include all values in block %s", b)
   308  		}
   309  	}
   310  
   311  	// Remove SPanchored now that we've scheduled.
   312  	// Also unlink nil checks now that ordering is assured
   313  	// between the nil check and the uses of the nil-checked pointer.
   314  	for _, b := range f.Blocks {
   315  		for _, v := range b.Values {
   316  			for i, a := range v.Args {
   317  				for a.Op == OpSPanchored || opcodeTable[a.Op].nilCheck {
   318  					a = a.Args[0]
   319  					v.SetArg(i, a)
   320  				}
   321  			}
   322  		}
   323  		for i, c := range b.ControlValues() {
   324  			for c.Op == OpSPanchored || opcodeTable[c.Op].nilCheck {
   325  				c = c.Args[0]
   326  				b.ReplaceControl(i, c)
   327  			}
   328  		}
   329  	}
   330  	for _, b := range f.Blocks {
   331  		i := 0
   332  		for _, v := range b.Values {
   333  			if v.Op == OpSPanchored {
   334  				// Free this value
   335  				if v.Uses != 0 {
   336  					base.Fatalf("SPAnchored still has %d uses", v.Uses)
   337  				}
   338  				v.resetArgs()
   339  				f.freeValue(v)
   340  			} else {
   341  				if opcodeTable[v.Op].nilCheck {
   342  					if v.Uses != 0 {
   343  						base.Fatalf("nilcheck still has %d uses", v.Uses)
   344  					}
   345  					// We can't delete the nil check, but we mark
   346  					// it as having void type so regalloc won't
   347  					// try to allocate a register for it.
   348  					v.Type = types.TypeVoid
   349  				}
   350  				b.Values[i] = v
   351  				i++
   352  			}
   353  		}
   354  		b.truncateValues(i)
   355  	}
   356  
   357  	f.scheduled = true
   358  }
   359  
   360  // storeOrder orders values with respect to stores. That is,
   361  // if v transitively depends on store s, v is ordered after s,
   362  // otherwise v is ordered before s.
   363  // Specifically, values are ordered like
   364  //
   365  //	store1
   366  //	NilCheck that depends on store1
   367  //	other values that depends on store1
   368  //	store2
   369  //	NilCheck that depends on store2
   370  //	other values that depends on store2
   371  //	...
   372  //
   373  // The order of non-store and non-NilCheck values are undefined
   374  // (not necessarily dependency order). This should be cheaper
   375  // than a full scheduling as done above.
   376  // Note that simple dependency order won't work: there is no
   377  // dependency between NilChecks and values like IsNonNil.
   378  // Auxiliary data structures are passed in as arguments, so
   379  // that they can be allocated in the caller and be reused.
   380  // This function takes care of reset them.
   381  func storeOrder(values []*Value, sset *sparseSet, storeNumber []int32) []*Value {
   382  	if len(values) == 0 {
   383  		return values
   384  	}
   385  
   386  	f := values[0].Block.Func
   387  
   388  	// find all stores
   389  
   390  	// Members of values that are store values.
   391  	// A constant bound allows this to be stack-allocated. 64 is
   392  	// enough to cover almost every storeOrder call.
   393  	stores := make([]*Value, 0, 64)
   394  	hasNilCheck := false
   395  	sset.clear() // sset is the set of stores that are used in other values
   396  	for _, v := range values {
   397  		if v.Type.IsMemory() {
   398  			stores = append(stores, v)
   399  			if v.Op == OpInitMem || v.Op == OpPhi {
   400  				continue
   401  			}
   402  			sset.add(v.MemoryArg().ID) // record that v's memory arg is used
   403  		}
   404  		if v.Op == OpNilCheck {
   405  			hasNilCheck = true
   406  		}
   407  	}
   408  	if len(stores) == 0 || !hasNilCheck && f.pass.name == "nilcheckelim" {
   409  		// there is no store, the order does not matter
   410  		return values
   411  	}
   412  
   413  	// find last store, which is the one that is not used by other stores
   414  	var last *Value
   415  	for _, v := range stores {
   416  		if !sset.contains(v.ID) {
   417  			if last != nil {
   418  				f.Fatalf("two stores live simultaneously: %v and %v", v, last)
   419  			}
   420  			last = v
   421  		}
   422  	}
   423  
   424  	// We assign a store number to each value. Store number is the
   425  	// index of the latest store that this value transitively depends.
   426  	// The i-th store in the current block gets store number 3*i. A nil
   427  	// check that depends on the i-th store gets store number 3*i+1.
   428  	// Other values that depends on the i-th store gets store number 3*i+2.
   429  	// Special case: 0 -- unassigned, 1 or 2 -- the latest store it depends
   430  	// is in the previous block (or no store at all, e.g. value is Const).
   431  	// First we assign the number to all stores by walking back the store chain,
   432  	// then assign the number to other values in DFS order.
   433  	count := make([]int32, 3*(len(stores)+1))
   434  	sset.clear() // reuse sparse set to ensure that a value is pushed to stack only once
   435  	for n, w := len(stores), last; n > 0; n-- {
   436  		storeNumber[w.ID] = int32(3 * n)
   437  		count[3*n]++
   438  		sset.add(w.ID)
   439  		if w.Op == OpInitMem || w.Op == OpPhi {
   440  			if n != 1 {
   441  				f.Fatalf("store order is wrong: there are stores before %v", w)
   442  			}
   443  			break
   444  		}
   445  		w = w.MemoryArg()
   446  	}
   447  	var stack []*Value
   448  	for _, v := range values {
   449  		if sset.contains(v.ID) {
   450  			// in sset means v is a store, or already pushed to stack, or already assigned a store number
   451  			continue
   452  		}
   453  		stack = append(stack, v)
   454  		sset.add(v.ID)
   455  
   456  		for len(stack) > 0 {
   457  			w := stack[len(stack)-1]
   458  			if storeNumber[w.ID] != 0 {
   459  				stack = stack[:len(stack)-1]
   460  				continue
   461  			}
   462  			if w.Op == OpPhi {
   463  				// Phi value doesn't depend on store in the current block.
   464  				// Do this early to avoid dependency cycle.
   465  				storeNumber[w.ID] = 2
   466  				count[2]++
   467  				stack = stack[:len(stack)-1]
   468  				continue
   469  			}
   470  
   471  			max := int32(0) // latest store dependency
   472  			argsdone := true
   473  			for _, a := range w.Args {
   474  				if a.Block != w.Block {
   475  					continue
   476  				}
   477  				if !sset.contains(a.ID) {
   478  					stack = append(stack, a)
   479  					sset.add(a.ID)
   480  					argsdone = false
   481  					break
   482  				}
   483  				if storeNumber[a.ID]/3 > max {
   484  					max = storeNumber[a.ID] / 3
   485  				}
   486  			}
   487  			if !argsdone {
   488  				continue
   489  			}
   490  
   491  			n := 3*max + 2
   492  			if w.Op == OpNilCheck {
   493  				n = 3*max + 1
   494  			}
   495  			storeNumber[w.ID] = n
   496  			count[n]++
   497  			stack = stack[:len(stack)-1]
   498  		}
   499  	}
   500  
   501  	// convert count to prefix sum of counts: count'[i] = sum_{j<=i} count[i]
   502  	for i := range count {
   503  		if i == 0 {
   504  			continue
   505  		}
   506  		count[i] += count[i-1]
   507  	}
   508  	if count[len(count)-1] != int32(len(values)) {
   509  		f.Fatalf("storeOrder: value is missing, total count = %d, values = %v", count[len(count)-1], values)
   510  	}
   511  
   512  	// place values in count-indexed bins, which are in the desired store order
   513  	order := make([]*Value, len(values))
   514  	for _, v := range values {
   515  		s := storeNumber[v.ID]
   516  		order[count[s-1]] = v
   517  		count[s-1]++
   518  	}
   519  
   520  	// Order nil checks in source order. We want the first in source order to trigger.
   521  	// If two are on the same line, we don't really care which happens first.
   522  	// See issue 18169.
   523  	if hasNilCheck {
   524  		start := -1
   525  		for i, v := range order {
   526  			if v.Op == OpNilCheck {
   527  				if start == -1 {
   528  					start = i
   529  				}
   530  			} else {
   531  				if start != -1 {
   532  					slices.SortFunc(order[start:i], valuePosCmp)
   533  					start = -1
   534  				}
   535  			}
   536  		}
   537  		if start != -1 {
   538  			slices.SortFunc(order[start:], valuePosCmp)
   539  		}
   540  	}
   541  
   542  	return order
   543  }
   544  
   545  // isFlagOp reports if v is an OP with the flag type.
   546  func (v *Value) isFlagOp() bool {
   547  	if v.Type.IsFlags() || v.Type.IsTuple() && v.Type.FieldType(1).IsFlags() {
   548  		return true
   549  	}
   550  	// PPC64 carry generators put their carry in a non-flag-typed register
   551  	// in their output.
   552  	switch v.Op {
   553  	case OpPPC64SUBC, OpPPC64ADDC, OpPPC64SUBCconst, OpPPC64ADDCconst:
   554  		return true
   555  	}
   556  	return false
   557  }
   558  
   559  // hasFlagInput reports whether v has a flag value as any of its inputs.
   560  func (v *Value) hasFlagInput() bool {
   561  	for _, a := range v.Args {
   562  		if a.isFlagOp() {
   563  			return true
   564  		}
   565  	}
   566  	// PPC64 carry dependencies are conveyed through their final argument,
   567  	// so we treat those operations as taking flags as well.
   568  	switch v.Op {
   569  	case OpPPC64SUBE, OpPPC64ADDE, OpPPC64SUBZEzero, OpPPC64ADDZE, OpPPC64ADDZEzero:
   570  		return true
   571  	}
   572  	return false
   573  }
   574  
   575  func valuePosCmp(a, b *Value) int {
   576  	if a.Pos.Before(b.Pos) {
   577  		return -1
   578  	}
   579  	if a.Pos.After(b.Pos) {
   580  		return +1
   581  	}
   582  	return 0
   583  }
   584
View as plain text