Source file src/cmd/compile/internal/ssa/rewrite.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 (
     8  	"cmd/compile/internal/logopt"
     9  	"cmd/compile/internal/types"
    10  	"cmd/internal/obj"
    11  	"cmd/internal/obj/s390x"
    12  	"cmd/internal/objabi"
    13  	"cmd/internal/src"
    14  	"encoding/binary"
    15  	"fmt"
    16  	"io"
    17  	"math"
    18  	"math/bits"
    19  	"os"
    20  	"path/filepath"
    21  )
    22  
    23  type deadValueChoice bool
    24  
    25  const (
    26  	leaveDeadValues  deadValueChoice = false
    27  	removeDeadValues                 = true
    28  )
    29  
    30  // deadcode indicates that rewrite should try to remove any values that become dead.
    31  func applyRewrite(f *Func, rb blockRewriter, rv valueRewriter, deadcode deadValueChoice) {
    32  	// repeat rewrites until we find no more rewrites
    33  	pendingLines := f.cachedLineStarts // Holds statement boundaries that need to be moved to a new value/block
    34  	pendingLines.clear()
    35  	debug := f.pass.debug
    36  	if debug > 1 {
    37  		fmt.Printf("%s: rewriting for %s\n", f.pass.name, f.Name)
    38  	}
    39  	for {
    40  		change := false
    41  		for _, b := range f.Blocks {
    42  			var b0 *Block
    43  			if debug > 1 {
    44  				b0 = new(Block)
    45  				*b0 = *b
    46  				b0.Succs = append([]Edge{}, b.Succs...) // make a new copy, not aliasing
    47  			}
    48  			for i, c := range b.ControlValues() {
    49  				for c.Op == OpCopy {
    50  					c = c.Args[0]
    51  					b.ReplaceControl(i, c)
    52  				}
    53  			}
    54  			if rb(b) {
    55  				change = true
    56  				if debug > 1 {
    57  					fmt.Printf("rewriting %s  ->  %s\n", b0.LongString(), b.LongString())
    58  				}
    59  			}
    60  			for j, v := range b.Values {
    61  				var v0 *Value
    62  				if debug > 1 {
    63  					v0 = new(Value)
    64  					*v0 = *v
    65  					v0.Args = append([]*Value{}, v.Args...) // make a new copy, not aliasing
    66  				}
    67  				if v.Uses == 0 && v.removeable() {
    68  					if v.Op != OpInvalid && deadcode == removeDeadValues {
    69  						// Reset any values that are now unused, so that we decrement
    70  						// the use count of all of its arguments.
    71  						// Not quite a deadcode pass, because it does not handle cycles.
    72  						// But it should help Uses==1 rules to fire.
    73  						v.reset(OpInvalid)
    74  						change = true
    75  					}
    76  					// No point rewriting values which aren't used.
    77  					continue
    78  				}
    79  
    80  				vchange := phielimValue(v)
    81  				if vchange && debug > 1 {
    82  					fmt.Printf("rewriting %s  ->  %s\n", v0.LongString(), v.LongString())
    83  				}
    84  
    85  				// Eliminate copy inputs.
    86  				// If any copy input becomes unused, mark it
    87  				// as invalid and discard its argument. Repeat
    88  				// recursively on the discarded argument.
    89  				// This phase helps remove phantom "dead copy" uses
    90  				// of a value so that a x.Uses==1 rule condition
    91  				// fires reliably.
    92  				for i, a := range v.Args {
    93  					if a.Op != OpCopy {
    94  						continue
    95  					}
    96  					aa := copySource(a)
    97  					v.SetArg(i, aa)
    98  					// If a, a copy, has a line boundary indicator, attempt to find a new value
    99  					// to hold it.  The first candidate is the value that will replace a (aa),
   100  					// if it shares the same block and line and is eligible.
   101  					// The second option is v, which has a as an input.  Because aa is earlier in
   102  					// the data flow, it is the better choice.
   103  					if a.Pos.IsStmt() == src.PosIsStmt {
   104  						if aa.Block == a.Block && aa.Pos.Line() == a.Pos.Line() && aa.Pos.IsStmt() != src.PosNotStmt {
   105  							aa.Pos = aa.Pos.WithIsStmt()
   106  						} else if v.Block == a.Block && v.Pos.Line() == a.Pos.Line() && v.Pos.IsStmt() != src.PosNotStmt {
   107  							v.Pos = v.Pos.WithIsStmt()
   108  						} else {
   109  							// Record the lost line and look for a new home after all rewrites are complete.
   110  							// TODO: it's possible (in FOR loops, in particular) for statement boundaries for the same
   111  							// line to appear in more than one block, but only one block is stored, so if both end
   112  							// up here, then one will be lost.
   113  							pendingLines.set(a.Pos, int32(a.Block.ID))
   114  						}
   115  						a.Pos = a.Pos.WithNotStmt()
   116  					}
   117  					vchange = true
   118  					for a.Uses == 0 {
   119  						b := a.Args[0]
   120  						a.reset(OpInvalid)
   121  						a = b
   122  					}
   123  				}
   124  				if vchange && debug > 1 {
   125  					fmt.Printf("rewriting %s  ->  %s\n", v0.LongString(), v.LongString())
   126  				}
   127  
   128  				// apply rewrite function
   129  				if rv(v) {
   130  					vchange = true
   131  					// If value changed to a poor choice for a statement boundary, move the boundary
   132  					if v.Pos.IsStmt() == src.PosIsStmt {
   133  						if k := nextGoodStatementIndex(v, j, b); k != j {
   134  							v.Pos = v.Pos.WithNotStmt()
   135  							b.Values[k].Pos = b.Values[k].Pos.WithIsStmt()
   136  						}
   137  					}
   138  				}
   139  
   140  				change = change || vchange
   141  				if vchange && debug > 1 {
   142  					fmt.Printf("rewriting %s  ->  %s\n", v0.LongString(), v.LongString())
   143  				}
   144  			}
   145  		}
   146  		if !change {
   147  			break
   148  		}
   149  	}
   150  	// remove clobbered values
   151  	for _, b := range f.Blocks {
   152  		j := 0
   153  		for i, v := range b.Values {
   154  			vl := v.Pos
   155  			if v.Op == OpInvalid {
   156  				if v.Pos.IsStmt() == src.PosIsStmt {
   157  					pendingLines.set(vl, int32(b.ID))
   158  				}
   159  				f.freeValue(v)
   160  				continue
   161  			}
   162  			if v.Pos.IsStmt() != src.PosNotStmt && pendingLines.get(vl) == int32(b.ID) {
   163  				pendingLines.remove(vl)
   164  				v.Pos = v.Pos.WithIsStmt()
   165  			}
   166  			if i != j {
   167  				b.Values[j] = v
   168  			}
   169  			j++
   170  		}
   171  		if pendingLines.get(b.Pos) == int32(b.ID) {
   172  			b.Pos = b.Pos.WithIsStmt()
   173  			pendingLines.remove(b.Pos)
   174  		}
   175  		b.truncateValues(j)
   176  	}
   177  }
   178  
   179  // Common functions called from rewriting rules
   180  
   181  func is64BitFloat(t *types.Type) bool {
   182  	return t.Size() == 8 && t.IsFloat()
   183  }
   184  
   185  func is32BitFloat(t *types.Type) bool {
   186  	return t.Size() == 4 && t.IsFloat()
   187  }
   188  
   189  func is64BitInt(t *types.Type) bool {
   190  	return t.Size() == 8 && t.IsInteger()
   191  }
   192  
   193  func is32BitInt(t *types.Type) bool {
   194  	return t.Size() == 4 && t.IsInteger()
   195  }
   196  
   197  func is16BitInt(t *types.Type) bool {
   198  	return t.Size() == 2 && t.IsInteger()
   199  }
   200  
   201  func is8BitInt(t *types.Type) bool {
   202  	return t.Size() == 1 && t.IsInteger()
   203  }
   204  
   205  func isPtr(t *types.Type) bool {
   206  	return t.IsPtrShaped()
   207  }
   208  
   209  func isSigned(t *types.Type) bool {
   210  	return t.IsSigned()
   211  }
   212  
   213  // mergeSym merges two symbolic offsets. There is no real merging of
   214  // offsets, we just pick the non-nil one.
   215  func mergeSym(x, y Sym) Sym {
   216  	if x == nil {
   217  		return y
   218  	}
   219  	if y == nil {
   220  		return x
   221  	}
   222  	panic(fmt.Sprintf("mergeSym with two non-nil syms %v %v", x, y))
   223  }
   224  
   225  func canMergeSym(x, y Sym) bool {
   226  	return x == nil || y == nil
   227  }
   228  
   229  // canMergeLoadClobber reports whether the load can be merged into target without
   230  // invalidating the schedule.
   231  // It also checks that the other non-load argument x is something we
   232  // are ok with clobbering.
   233  func canMergeLoadClobber(target, load, x *Value) bool {
   234  	// The register containing x is going to get clobbered.
   235  	// Don't merge if we still need the value of x.
   236  	// We don't have liveness information here, but we can
   237  	// approximate x dying with:
   238  	//  1) target is x's only use.
   239  	//  2) target is not in a deeper loop than x.
   240  	if x.Uses != 1 {
   241  		return false
   242  	}
   243  	loopnest := x.Block.Func.loopnest()
   244  	loopnest.calculateDepths()
   245  	if loopnest.depth(target.Block.ID) > loopnest.depth(x.Block.ID) {
   246  		return false
   247  	}
   248  	return canMergeLoad(target, load)
   249  }
   250  
   251  // canMergeLoad reports whether the load can be merged into target without
   252  // invalidating the schedule.
   253  func canMergeLoad(target, load *Value) bool {
   254  	if target.Block.ID != load.Block.ID {
   255  		// If the load is in a different block do not merge it.
   256  		return false
   257  	}
   258  
   259  	// We can't merge the load into the target if the load
   260  	// has more than one use.
   261  	if load.Uses != 1 {
   262  		return false
   263  	}
   264  
   265  	mem := load.MemoryArg()
   266  
   267  	// We need the load's memory arg to still be alive at target. That
   268  	// can't be the case if one of target's args depends on a memory
   269  	// state that is a successor of load's memory arg.
   270  	//
   271  	// For example, it would be invalid to merge load into target in
   272  	// the following situation because newmem has killed oldmem
   273  	// before target is reached:
   274  	//     load = read ... oldmem
   275  	//   newmem = write ... oldmem
   276  	//     arg0 = read ... newmem
   277  	//   target = add arg0 load
   278  	//
   279  	// If the argument comes from a different block then we can exclude
   280  	// it immediately because it must dominate load (which is in the
   281  	// same block as target).
   282  	var args []*Value
   283  	for _, a := range target.Args {
   284  		if a != load && a.Block.ID == target.Block.ID {
   285  			args = append(args, a)
   286  		}
   287  	}
   288  
   289  	// memPreds contains memory states known to be predecessors of load's
   290  	// memory state. It is lazily initialized.
   291  	var memPreds map[*Value]bool
   292  	for i := 0; len(args) > 0; i++ {
   293  		const limit = 100
   294  		if i >= limit {
   295  			// Give up if we have done a lot of iterations.
   296  			return false
   297  		}
   298  		v := args[len(args)-1]
   299  		args = args[:len(args)-1]
   300  		if target.Block.ID != v.Block.ID {
   301  			// Since target and load are in the same block
   302  			// we can stop searching when we leave the block.
   303  			continue
   304  		}
   305  		if v.Op == OpPhi {
   306  			// A Phi implies we have reached the top of the block.
   307  			// The memory phi, if it exists, is always
   308  			// the first logical store in the block.
   309  			continue
   310  		}
   311  		if v.Type.IsTuple() && v.Type.FieldType(1).IsMemory() {
   312  			// We could handle this situation however it is likely
   313  			// to be very rare.
   314  			return false
   315  		}
   316  		if v.Op.SymEffect()&SymAddr != 0 {
   317  			// This case prevents an operation that calculates the
   318  			// address of a local variable from being forced to schedule
   319  			// before its corresponding VarDef.
   320  			// See issue 28445.
   321  			//   v1 = LOAD ...
   322  			//   v2 = VARDEF
   323  			//   v3 = LEAQ
   324  			//   v4 = CMPQ v1 v3
   325  			// We don't want to combine the CMPQ with the load, because
   326  			// that would force the CMPQ to schedule before the VARDEF, which
   327  			// in turn requires the LEAQ to schedule before the VARDEF.
   328  			return false
   329  		}
   330  		if v.Type.IsMemory() {
   331  			if memPreds == nil {
   332  				// Initialise a map containing memory states
   333  				// known to be predecessors of load's memory
   334  				// state.
   335  				memPreds = make(map[*Value]bool)
   336  				m := mem
   337  				const limit = 50
   338  				for i := 0; i < limit; i++ {
   339  					if m.Op == OpPhi {
   340  						// The memory phi, if it exists, is always
   341  						// the first logical store in the block.
   342  						break
   343  					}
   344  					if m.Block.ID != target.Block.ID {
   345  						break
   346  					}
   347  					if !m.Type.IsMemory() {
   348  						break
   349  					}
   350  					memPreds[m] = true
   351  					if len(m.Args) == 0 {
   352  						break
   353  					}
   354  					m = m.MemoryArg()
   355  				}
   356  			}
   357  
   358  			// We can merge if v is a predecessor of mem.
   359  			//
   360  			// For example, we can merge load into target in the
   361  			// following scenario:
   362  			//      x = read ... v
   363  			//    mem = write ... v
   364  			//   load = read ... mem
   365  			// target = add x load
   366  			if memPreds[v] {
   367  				continue
   368  			}
   369  			return false
   370  		}
   371  		if len(v.Args) > 0 && v.Args[len(v.Args)-1] == mem {
   372  			// If v takes mem as an input then we know mem
   373  			// is valid at this point.
   374  			continue
   375  		}
   376  		for _, a := range v.Args {
   377  			if target.Block.ID == a.Block.ID {
   378  				args = append(args, a)
   379  			}
   380  		}
   381  	}
   382  
   383  	return true
   384  }
   385  
   386  // isSameCall reports whether sym is the same as the given named symbol
   387  func isSameCall(sym interface{}, name string) bool {
   388  	fn := sym.(*AuxCall).Fn
   389  	return fn != nil && fn.String() == name
   390  }
   391  
   392  // nlz returns the number of leading zeros.
   393  func nlz64(x int64) int { return bits.LeadingZeros64(uint64(x)) }
   394  func nlz32(x int32) int { return bits.LeadingZeros32(uint32(x)) }
   395  func nlz16(x int16) int { return bits.LeadingZeros16(uint16(x)) }
   396  func nlz8(x int8) int   { return bits.LeadingZeros8(uint8(x)) }
   397  
   398  // ntzX returns the number of trailing zeros.
   399  func ntz64(x int64) int { return bits.TrailingZeros64(uint64(x)) }
   400  func ntz32(x int32) int { return bits.TrailingZeros32(uint32(x)) }
   401  func ntz16(x int16) int { return bits.TrailingZeros16(uint16(x)) }
   402  func ntz8(x int8) int   { return bits.TrailingZeros8(uint8(x)) }
   403  
   404  func oneBit(x int64) bool   { return x&(x-1) == 0 && x != 0 }
   405  func oneBit8(x int8) bool   { return x&(x-1) == 0 && x != 0 }
   406  func oneBit16(x int16) bool { return x&(x-1) == 0 && x != 0 }
   407  func oneBit32(x int32) bool { return x&(x-1) == 0 && x != 0 }
   408  func oneBit64(x int64) bool { return x&(x-1) == 0 && x != 0 }
   409  
   410  // nto returns the number of trailing ones.
   411  func nto(x int64) int64 {
   412  	return int64(ntz64(^x))
   413  }
   414  
   415  // logX returns logarithm of n base 2.
   416  // n must be a positive power of 2 (isPowerOfTwoX returns true).
   417  func log8(n int8) int64 {
   418  	return int64(bits.Len8(uint8(n))) - 1
   419  }
   420  func log16(n int16) int64 {
   421  	return int64(bits.Len16(uint16(n))) - 1
   422  }
   423  func log32(n int32) int64 {
   424  	return int64(bits.Len32(uint32(n))) - 1
   425  }
   426  func log64(n int64) int64 {
   427  	return int64(bits.Len64(uint64(n))) - 1
   428  }
   429  
   430  // log2uint32 returns logarithm in base 2 of uint32(n), with log2(0) = -1.
   431  // Rounds down.
   432  func log2uint32(n int64) int64 {
   433  	return int64(bits.Len32(uint32(n))) - 1
   434  }
   435  
   436  // isPowerOfTwo functions report whether n is a power of 2.
   437  func isPowerOfTwo8(n int8) bool {
   438  	return n > 0 && n&(n-1) == 0
   439  }
   440  func isPowerOfTwo16(n int16) bool {
   441  	return n > 0 && n&(n-1) == 0
   442  }
   443  func isPowerOfTwo32(n int32) bool {
   444  	return n > 0 && n&(n-1) == 0
   445  }
   446  func isPowerOfTwo64(n int64) bool {
   447  	return n > 0 && n&(n-1) == 0
   448  }
   449  
   450  // isUint64PowerOfTwo reports whether uint64(n) is a power of 2.
   451  func isUint64PowerOfTwo(in int64) bool {
   452  	n := uint64(in)
   453  	return n > 0 && n&(n-1) == 0
   454  }
   455  
   456  // isUint32PowerOfTwo reports whether uint32(n) is a power of 2.
   457  func isUint32PowerOfTwo(in int64) bool {
   458  	n := uint64(uint32(in))
   459  	return n > 0 && n&(n-1) == 0
   460  }
   461  
   462  // is32Bit reports whether n can be represented as a signed 32 bit integer.
   463  func is32Bit(n int64) bool {
   464  	return n == int64(int32(n))
   465  }
   466  
   467  // is16Bit reports whether n can be represented as a signed 16 bit integer.
   468  func is16Bit(n int64) bool {
   469  	return n == int64(int16(n))
   470  }
   471  
   472  // is8Bit reports whether n can be represented as a signed 8 bit integer.
   473  func is8Bit(n int64) bool {
   474  	return n == int64(int8(n))
   475  }
   476  
   477  // isU8Bit reports whether n can be represented as an unsigned 8 bit integer.
   478  func isU8Bit(n int64) bool {
   479  	return n == int64(uint8(n))
   480  }
   481  
   482  // isU12Bit reports whether n can be represented as an unsigned 12 bit integer.
   483  func isU12Bit(n int64) bool {
   484  	return 0 <= n && n < (1<<12)
   485  }
   486  
   487  // isU16Bit reports whether n can be represented as an unsigned 16 bit integer.
   488  func isU16Bit(n int64) bool {
   489  	return n == int64(uint16(n))
   490  }
   491  
   492  // isU32Bit reports whether n can be represented as an unsigned 32 bit integer.
   493  func isU32Bit(n int64) bool {
   494  	return n == int64(uint32(n))
   495  }
   496  
   497  // is20Bit reports whether n can be represented as a signed 20 bit integer.
   498  func is20Bit(n int64) bool {
   499  	return -(1<<19) <= n && n < (1<<19)
   500  }
   501  
   502  // b2i translates a boolean value to 0 or 1 for assigning to auxInt.
   503  func b2i(b bool) int64 {
   504  	if b {
   505  		return 1
   506  	}
   507  	return 0
   508  }
   509  
   510  // b2i32 translates a boolean value to 0 or 1.
   511  func b2i32(b bool) int32 {
   512  	if b {
   513  		return 1
   514  	}
   515  	return 0
   516  }
   517  
   518  // shiftIsBounded reports whether (left/right) shift Value v is known to be bounded.
   519  // A shift is bounded if it is shifting by less than the width of the shifted value.
   520  func shiftIsBounded(v *Value) bool {
   521  	return v.AuxInt != 0
   522  }
   523  
   524  // truncate64Fto32F converts a float64 value to a float32 preserving the bit pattern
   525  // of the mantissa. It will panic if the truncation results in lost information.
   526  func truncate64Fto32F(f float64) float32 {
   527  	if !isExactFloat32(f) {
   528  		panic("truncate64Fto32F: truncation is not exact")
   529  	}
   530  	if !math.IsNaN(f) {
   531  		return float32(f)
   532  	}
   533  	// NaN bit patterns aren't necessarily preserved across conversion
   534  	// instructions so we need to do the conversion manually.
   535  	b := math.Float64bits(f)
   536  	m := b & ((1 << 52) - 1) // mantissa (a.k.a. significand)
   537  	//          | sign                  | exponent   | mantissa       |
   538  	r := uint32(((b >> 32) & (1 << 31)) | 0x7f800000 | (m >> (52 - 23)))
   539  	return math.Float32frombits(r)
   540  }
   541  
   542  // extend32Fto64F converts a float32 value to a float64 value preserving the bit
   543  // pattern of the mantissa.
   544  func extend32Fto64F(f float32) float64 {
   545  	if !math.IsNaN(float64(f)) {
   546  		return float64(f)
   547  	}
   548  	// NaN bit patterns aren't necessarily preserved across conversion
   549  	// instructions so we need to do the conversion manually.
   550  	b := uint64(math.Float32bits(f))
   551  	//   | sign                  | exponent      | mantissa                    |
   552  	r := ((b << 32) & (1 << 63)) | (0x7ff << 52) | ((b & 0x7fffff) << (52 - 23))
   553  	return math.Float64frombits(r)
   554  }
   555  
   556  // DivisionNeedsFixUp reports whether the division needs fix-up code.
   557  func DivisionNeedsFixUp(v *Value) bool {
   558  	return v.AuxInt == 0
   559  }
   560  
   561  // auxFrom64F encodes a float64 value so it can be stored in an AuxInt.
   562  func auxFrom64F(f float64) int64 {
   563  	if f != f {
   564  		panic("can't encode a NaN in AuxInt field")
   565  	}
   566  	return int64(math.Float64bits(f))
   567  }
   568  
   569  // auxFrom32F encodes a float32 value so it can be stored in an AuxInt.
   570  func auxFrom32F(f float32) int64 {
   571  	if f != f {
   572  		panic("can't encode a NaN in AuxInt field")
   573  	}
   574  	return int64(math.Float64bits(extend32Fto64F(f)))
   575  }
   576  
   577  // auxTo32F decodes a float32 from the AuxInt value provided.
   578  func auxTo32F(i int64) float32 {
   579  	return truncate64Fto32F(math.Float64frombits(uint64(i)))
   580  }
   581  
   582  // auxTo64F decodes a float64 from the AuxInt value provided.
   583  func auxTo64F(i int64) float64 {
   584  	return math.Float64frombits(uint64(i))
   585  }
   586  
   587  func auxIntToBool(i int64) bool {
   588  	if i == 0 {
   589  		return false
   590  	}
   591  	return true
   592  }
   593  func auxIntToInt8(i int64) int8 {
   594  	return int8(i)
   595  }
   596  func auxIntToInt16(i int64) int16 {
   597  	return int16(i)
   598  }
   599  func auxIntToInt32(i int64) int32 {
   600  	return int32(i)
   601  }
   602  func auxIntToInt64(i int64) int64 {
   603  	return i
   604  }
   605  func auxIntToUint8(i int64) uint8 {
   606  	return uint8(i)
   607  }
   608  func auxIntToFloat32(i int64) float32 {
   609  	return float32(math.Float64frombits(uint64(i)))
   610  }
   611  func auxIntToFloat64(i int64) float64 {
   612  	return math.Float64frombits(uint64(i))
   613  }
   614  func auxIntToValAndOff(i int64) ValAndOff {
   615  	return ValAndOff(i)
   616  }
   617  func auxIntToArm64BitField(i int64) arm64BitField {
   618  	return arm64BitField(i)
   619  }
   620  func auxIntToInt128(x int64) int128 {
   621  	if x != 0 {
   622  		panic("nonzero int128 not allowed")
   623  	}
   624  	return 0
   625  }
   626  func auxIntToFlagConstant(x int64) flagConstant {
   627  	return flagConstant(x)
   628  }
   629  
   630  func auxIntToOp(cc int64) Op {
   631  	return Op(cc)
   632  }
   633  
   634  func boolToAuxInt(b bool) int64 {
   635  	if b {
   636  		return 1
   637  	}
   638  	return 0
   639  }
   640  func int8ToAuxInt(i int8) int64 {
   641  	return int64(i)
   642  }
   643  func int16ToAuxInt(i int16) int64 {
   644  	return int64(i)
   645  }
   646  func int32ToAuxInt(i int32) int64 {
   647  	return int64(i)
   648  }
   649  func int64ToAuxInt(i int64) int64 {
   650  	return int64(i)
   651  }
   652  func uint8ToAuxInt(i uint8) int64 {
   653  	return int64(int8(i))
   654  }
   655  func float32ToAuxInt(f float32) int64 {
   656  	return int64(math.Float64bits(float64(f)))
   657  }
   658  func float64ToAuxInt(f float64) int64 {
   659  	return int64(math.Float64bits(f))
   660  }
   661  func valAndOffToAuxInt(v ValAndOff) int64 {
   662  	return int64(v)
   663  }
   664  func arm64BitFieldToAuxInt(v arm64BitField) int64 {
   665  	return int64(v)
   666  }
   667  func int128ToAuxInt(x int128) int64 {
   668  	if x != 0 {
   669  		panic("nonzero int128 not allowed")
   670  	}
   671  	return 0
   672  }
   673  func flagConstantToAuxInt(x flagConstant) int64 {
   674  	return int64(x)
   675  }
   676  
   677  func opToAuxInt(o Op) int64 {
   678  	return int64(o)
   679  }
   680  
   681  func auxToString(i interface{}) string {
   682  	return i.(string)
   683  }
   684  func auxToSym(i interface{}) Sym {
   685  	// TODO: kind of a hack - allows nil interface through
   686  	s, _ := i.(Sym)
   687  	return s
   688  }
   689  func auxToType(i interface{}) *types.Type {
   690  	return i.(*types.Type)
   691  }
   692  func auxToCall(i interface{}) *AuxCall {
   693  	return i.(*AuxCall)
   694  }
   695  func auxToS390xCCMask(i interface{}) s390x.CCMask {
   696  	return i.(s390x.CCMask)
   697  }
   698  func auxToS390xRotateParams(i interface{}) s390x.RotateParams {
   699  	return i.(s390x.RotateParams)
   700  }
   701  
   702  func stringToAux(s string) interface{} {
   703  	return s
   704  }
   705  func symToAux(s Sym) interface{} {
   706  	return s
   707  }
   708  func callToAux(s *AuxCall) interface{} {
   709  	return s
   710  }
   711  func typeToAux(t *types.Type) interface{} {
   712  	return t
   713  }
   714  func s390xCCMaskToAux(c s390x.CCMask) interface{} {
   715  	return c
   716  }
   717  func s390xRotateParamsToAux(r s390x.RotateParams) interface{} {
   718  	return r
   719  }
   720  
   721  // uaddOvf reports whether unsigned a+b would overflow.
   722  func uaddOvf(a, b int64) bool {
   723  	return uint64(a)+uint64(b) < uint64(a)
   724  }
   725  
   726  // de-virtualize an InterCall
   727  // 'sym' is the symbol for the itab
   728  func devirt(v *Value, aux interface{}, sym Sym, offset int64) *AuxCall {
   729  	f := v.Block.Func
   730  	n, ok := sym.(*obj.LSym)
   731  	if !ok {
   732  		return nil
   733  	}
   734  	lsym := f.fe.DerefItab(n, offset)
   735  	if f.pass.debug > 0 {
   736  		if lsym != nil {
   737  			f.Warnl(v.Pos, "de-virtualizing call")
   738  		} else {
   739  			f.Warnl(v.Pos, "couldn't de-virtualize call")
   740  		}
   741  	}
   742  	if lsym == nil {
   743  		return nil
   744  	}
   745  	va := aux.(*AuxCall)
   746  	return StaticAuxCall(lsym, va.args, va.results)
   747  }
   748  
   749  // de-virtualize an InterLECall
   750  // 'sym' is the symbol for the itab
   751  func devirtLESym(v *Value, aux interface{}, sym Sym, offset int64) *obj.LSym {
   752  	n, ok := sym.(*obj.LSym)
   753  	if !ok {
   754  		return nil
   755  	}
   756  
   757  	f := v.Block.Func
   758  	lsym := f.fe.DerefItab(n, offset)
   759  	if f.pass.debug > 0 {
   760  		if lsym != nil {
   761  			f.Warnl(v.Pos, "de-virtualizing call")
   762  		} else {
   763  			f.Warnl(v.Pos, "couldn't de-virtualize call")
   764  		}
   765  	}
   766  	if lsym == nil {
   767  		return nil
   768  	}
   769  	return lsym
   770  }
   771  
   772  func devirtLECall(v *Value, sym *obj.LSym) *Value {
   773  	v.Op = OpStaticLECall
   774  	v.Aux.(*AuxCall).Fn = sym
   775  	v.RemoveArg(0)
   776  	return v
   777  }
   778  
   779  // isSamePtr reports whether p1 and p2 point to the same address.
   780  func isSamePtr(p1, p2 *Value) bool {
   781  	if p1 == p2 {
   782  		return true
   783  	}
   784  	if p1.Op != p2.Op {
   785  		return false
   786  	}
   787  	switch p1.Op {
   788  	case OpOffPtr:
   789  		return p1.AuxInt == p2.AuxInt && isSamePtr(p1.Args[0], p2.Args[0])
   790  	case OpAddr, OpLocalAddr:
   791  		// OpAddr's 0th arg is either OpSP or OpSB, which means that it is uniquely identified by its Op.
   792  		// Checking for value equality only works after [z]cse has run.
   793  		return p1.Aux == p2.Aux && p1.Args[0].Op == p2.Args[0].Op
   794  	case OpAddPtr:
   795  		return p1.Args[1] == p2.Args[1] && isSamePtr(p1.Args[0], p2.Args[0])
   796  	}
   797  	return false
   798  }
   799  
   800  func isStackPtr(v *Value) bool {
   801  	for v.Op == OpOffPtr || v.Op == OpAddPtr {
   802  		v = v.Args[0]
   803  	}
   804  	return v.Op == OpSP || v.Op == OpLocalAddr
   805  }
   806  
   807  // disjoint reports whether the memory region specified by [p1:p1+n1)
   808  // does not overlap with [p2:p2+n2).
   809  // A return value of false does not imply the regions overlap.
   810  func disjoint(p1 *Value, n1 int64, p2 *Value, n2 int64) bool {
   811  	if n1 == 0 || n2 == 0 {
   812  		return true
   813  	}
   814  	if p1 == p2 {
   815  		return false
   816  	}
   817  	baseAndOffset := func(ptr *Value) (base *Value, offset int64) {
   818  		base, offset = ptr, 0
   819  		for base.Op == OpOffPtr {
   820  			offset += base.AuxInt
   821  			base = base.Args[0]
   822  		}
   823  		return base, offset
   824  	}
   825  	p1, off1 := baseAndOffset(p1)
   826  	p2, off2 := baseAndOffset(p2)
   827  	if isSamePtr(p1, p2) {
   828  		return !overlap(off1, n1, off2, n2)
   829  	}
   830  	// p1 and p2 are not the same, so if they are both OpAddrs then
   831  	// they point to different variables.
   832  	// If one pointer is on the stack and the other is an argument
   833  	// then they can't overlap.
   834  	switch p1.Op {
   835  	case OpAddr, OpLocalAddr:
   836  		if p2.Op == OpAddr || p2.Op == OpLocalAddr || p2.Op == OpSP {
   837  			return true
   838  		}
   839  		return p2.Op == OpArg && p1.Args[0].Op == OpSP
   840  	case OpArg:
   841  		if p2.Op == OpSP || p2.Op == OpLocalAddr {
   842  			return true
   843  		}
   844  	case OpSP:
   845  		return p2.Op == OpAddr || p2.Op == OpLocalAddr || p2.Op == OpArg || p2.Op == OpSP
   846  	}
   847  	return false
   848  }
   849  
   850  // moveSize returns the number of bytes an aligned MOV instruction moves
   851  func moveSize(align int64, c *Config) int64 {
   852  	switch {
   853  	case align%8 == 0 && c.PtrSize == 8:
   854  		return 8
   855  	case align%4 == 0:
   856  		return 4
   857  	case align%2 == 0:
   858  		return 2
   859  	}
   860  	return 1
   861  }
   862  
   863  // mergePoint finds a block among a's blocks which dominates b and is itself
   864  // dominated by all of a's blocks. Returns nil if it can't find one.
   865  // Might return nil even if one does exist.
   866  func mergePoint(b *Block, a ...*Value) *Block {
   867  	// Walk backward from b looking for one of the a's blocks.
   868  
   869  	// Max distance
   870  	d := 100
   871  
   872  	for d > 0 {
   873  		for _, x := range a {
   874  			if b == x.Block {
   875  				goto found
   876  			}
   877  		}
   878  		if len(b.Preds) > 1 {
   879  			// Don't know which way to go back. Abort.
   880  			return nil
   881  		}
   882  		b = b.Preds[0].b
   883  		d--
   884  	}
   885  	return nil // too far away
   886  found:
   887  	// At this point, r is the first value in a that we find by walking backwards.
   888  	// if we return anything, r will be it.
   889  	r := b
   890  
   891  	// Keep going, counting the other a's that we find. They must all dominate r.
   892  	na := 0
   893  	for d > 0 {
   894  		for _, x := range a {
   895  			if b == x.Block {
   896  				na++
   897  			}
   898  		}
   899  		if na == len(a) {
   900  			// Found all of a in a backwards walk. We can return r.
   901  			return r
   902  		}
   903  		if len(b.Preds) > 1 {
   904  			return nil
   905  		}
   906  		b = b.Preds[0].b
   907  		d--
   908  
   909  	}
   910  	return nil // too far away
   911  }
   912  
   913  // clobber invalidates values. Returns true.
   914  // clobber is used by rewrite rules to:
   915  //   A) make sure the values are really dead and never used again.
   916  //   B) decrement use counts of the values' args.
   917  func clobber(vv ...*Value) bool {
   918  	for _, v := range vv {
   919  		v.reset(OpInvalid)
   920  		// Note: leave v.Block intact.  The Block field is used after clobber.
   921  	}
   922  	return true
   923  }
   924  
   925  // clobberIfDead resets v when use count is 1. Returns true.
   926  // clobberIfDead is used by rewrite rules to decrement
   927  // use counts of v's args when v is dead and never used.
   928  func clobberIfDead(v *Value) bool {
   929  	if v.Uses == 1 {
   930  		v.reset(OpInvalid)
   931  	}
   932  	// Note: leave v.Block intact.  The Block field is used after clobberIfDead.
   933  	return true
   934  }
   935  
   936  // noteRule is an easy way to track if a rule is matched when writing
   937  // new ones.  Make the rule of interest also conditional on
   938  //     noteRule("note to self: rule of interest matched")
   939  // and that message will print when the rule matches.
   940  func noteRule(s string) bool {
   941  	fmt.Println(s)
   942  	return true
   943  }
   944  
   945  // countRule increments Func.ruleMatches[key].
   946  // If Func.ruleMatches is non-nil at the end
   947  // of compilation, it will be printed to stdout.
   948  // This is intended to make it easier to find which functions
   949  // which contain lots of rules matches when developing new rules.
   950  func countRule(v *Value, key string) bool {
   951  	f := v.Block.Func
   952  	if f.ruleMatches == nil {
   953  		f.ruleMatches = make(map[string]int)
   954  	}
   955  	f.ruleMatches[key]++
   956  	return true
   957  }
   958  
   959  // warnRule generates compiler debug output with string s when
   960  // v is not in autogenerated code, cond is true and the rule has fired.
   961  func warnRule(cond bool, v *Value, s string) bool {
   962  	if pos := v.Pos; pos.Line() > 1 && cond {
   963  		v.Block.Func.Warnl(pos, s)
   964  	}
   965  	return true
   966  }
   967  
   968  // for a pseudo-op like (LessThan x), extract x
   969  func flagArg(v *Value) *Value {
   970  	if len(v.Args) != 1 || !v.Args[0].Type.IsFlags() {
   971  		return nil
   972  	}
   973  	return v.Args[0]
   974  }
   975  
   976  // arm64Negate finds the complement to an ARM64 condition code,
   977  // for example !Equal -> NotEqual or !LessThan -> GreaterEqual
   978  //
   979  // For floating point, it's more subtle because NaN is unordered. We do
   980  // !LessThanF -> NotLessThanF, the latter takes care of NaNs.
   981  func arm64Negate(op Op) Op {
   982  	switch op {
   983  	case OpARM64LessThan:
   984  		return OpARM64GreaterEqual
   985  	case OpARM64LessThanU:
   986  		return OpARM64GreaterEqualU
   987  	case OpARM64GreaterThan:
   988  		return OpARM64LessEqual
   989  	case OpARM64GreaterThanU:
   990  		return OpARM64LessEqualU
   991  	case OpARM64LessEqual:
   992  		return OpARM64GreaterThan
   993  	case OpARM64LessEqualU:
   994  		return OpARM64GreaterThanU
   995  	case OpARM64GreaterEqual:
   996  		return OpARM64LessThan
   997  	case OpARM64GreaterEqualU:
   998  		return OpARM64LessThanU
   999  	case OpARM64Equal:
  1000  		return OpARM64NotEqual
  1001  	case OpARM64NotEqual:
  1002  		return OpARM64Equal
  1003  	case OpARM64LessThanF:
  1004  		return OpARM64NotLessThanF
  1005  	case OpARM64NotLessThanF:
  1006  		return OpARM64LessThanF
  1007  	case OpARM64LessEqualF:
  1008  		return OpARM64NotLessEqualF
  1009  	case OpARM64NotLessEqualF:
  1010  		return OpARM64LessEqualF
  1011  	case OpARM64GreaterThanF:
  1012  		return OpARM64NotGreaterThanF
  1013  	case OpARM64NotGreaterThanF:
  1014  		return OpARM64GreaterThanF
  1015  	case OpARM64GreaterEqualF:
  1016  		return OpARM64NotGreaterEqualF
  1017  	case OpARM64NotGreaterEqualF:
  1018  		return OpARM64GreaterEqualF
  1019  	default:
  1020  		panic("unreachable")
  1021  	}
  1022  }
  1023  
  1024  // arm64Invert evaluates (InvertFlags op), which
  1025  // is the same as altering the condition codes such
  1026  // that the same result would be produced if the arguments
  1027  // to the flag-generating instruction were reversed, e.g.
  1028  // (InvertFlags (CMP x y)) -> (CMP y x)
  1029  func arm64Invert(op Op) Op {
  1030  	switch op {
  1031  	case OpARM64LessThan:
  1032  		return OpARM64GreaterThan
  1033  	case OpARM64LessThanU:
  1034  		return OpARM64GreaterThanU
  1035  	case OpARM64GreaterThan:
  1036  		return OpARM64LessThan
  1037  	case OpARM64GreaterThanU:
  1038  		return OpARM64LessThanU
  1039  	case OpARM64LessEqual:
  1040  		return OpARM64GreaterEqual
  1041  	case OpARM64LessEqualU:
  1042  		return OpARM64GreaterEqualU
  1043  	case OpARM64GreaterEqual:
  1044  		return OpARM64LessEqual
  1045  	case OpARM64GreaterEqualU:
  1046  		return OpARM64LessEqualU
  1047  	case OpARM64Equal, OpARM64NotEqual:
  1048  		return op
  1049  	case OpARM64LessThanF:
  1050  		return OpARM64GreaterThanF
  1051  	case OpARM64GreaterThanF:
  1052  		return OpARM64LessThanF
  1053  	case OpARM64LessEqualF:
  1054  		return OpARM64GreaterEqualF
  1055  	case OpARM64GreaterEqualF:
  1056  		return OpARM64LessEqualF
  1057  	case OpARM64NotLessThanF:
  1058  		return OpARM64NotGreaterThanF
  1059  	case OpARM64NotGreaterThanF:
  1060  		return OpARM64NotLessThanF
  1061  	case OpARM64NotLessEqualF:
  1062  		return OpARM64NotGreaterEqualF
  1063  	case OpARM64NotGreaterEqualF:
  1064  		return OpARM64NotLessEqualF
  1065  	default:
  1066  		panic("unreachable")
  1067  	}
  1068  }
  1069  
  1070  // evaluate an ARM64 op against a flags value
  1071  // that is potentially constant; return 1 for true,
  1072  // -1 for false, and 0 for not constant.
  1073  func ccARM64Eval(op Op, flags *Value) int {
  1074  	fop := flags.Op
  1075  	if fop == OpARM64InvertFlags {
  1076  		return -ccARM64Eval(op, flags.Args[0])
  1077  	}
  1078  	if fop != OpARM64FlagConstant {
  1079  		return 0
  1080  	}
  1081  	fc := flagConstant(flags.AuxInt)
  1082  	b2i := func(b bool) int {
  1083  		if b {
  1084  			return 1
  1085  		}
  1086  		return -1
  1087  	}
  1088  	switch op {
  1089  	case OpARM64Equal:
  1090  		return b2i(fc.eq())
  1091  	case OpARM64NotEqual:
  1092  		return b2i(fc.ne())
  1093  	case OpARM64LessThan:
  1094  		return b2i(fc.lt())
  1095  	case OpARM64LessThanU:
  1096  		return b2i(fc.ult())
  1097  	case OpARM64GreaterThan:
  1098  		return b2i(fc.gt())
  1099  	case OpARM64GreaterThanU:
  1100  		return b2i(fc.ugt())
  1101  	case OpARM64LessEqual:
  1102  		return b2i(fc.le())
  1103  	case OpARM64LessEqualU:
  1104  		return b2i(fc.ule())
  1105  	case OpARM64GreaterEqual:
  1106  		return b2i(fc.ge())
  1107  	case OpARM64GreaterEqualU:
  1108  		return b2i(fc.uge())
  1109  	}
  1110  	return 0
  1111  }
  1112  
  1113  // logRule logs the use of the rule s. This will only be enabled if
  1114  // rewrite rules were generated with the -log option, see gen/rulegen.go.
  1115  func logRule(s string) {
  1116  	if ruleFile == nil {
  1117  		// Open a log file to write log to. We open in append
  1118  		// mode because all.bash runs the compiler lots of times,
  1119  		// and we want the concatenation of all of those logs.
  1120  		// This means, of course, that users need to rm the old log
  1121  		// to get fresh data.
  1122  		// TODO: all.bash runs compilers in parallel. Need to synchronize logging somehow?
  1123  		w, err := os.OpenFile(filepath.Join(os.Getenv("GOROOT"), "src", "rulelog"),
  1124  			os.O_CREATE|os.O_WRONLY|os.O_APPEND, 0666)
  1125  		if err != nil {
  1126  			panic(err)
  1127  		}
  1128  		ruleFile = w
  1129  	}
  1130  	_, err := fmt.Fprintln(ruleFile, s)
  1131  	if err != nil {
  1132  		panic(err)
  1133  	}
  1134  }
  1135  
  1136  var ruleFile io.Writer
  1137  
  1138  func min(x, y int64) int64 {
  1139  	if x < y {
  1140  		return x
  1141  	}
  1142  	return y
  1143  }
  1144  
  1145  func isConstZero(v *Value) bool {
  1146  	switch v.Op {
  1147  	case OpConstNil:
  1148  		return true
  1149  	case OpConst64, OpConst32, OpConst16, OpConst8, OpConstBool, OpConst32F, OpConst64F:
  1150  		return v.AuxInt == 0
  1151  	}
  1152  	return false
  1153  }
  1154  
  1155  // reciprocalExact64 reports whether 1/c is exactly representable.
  1156  func reciprocalExact64(c float64) bool {
  1157  	b := math.Float64bits(c)
  1158  	man := b & (1<<52 - 1)
  1159  	if man != 0 {
  1160  		return false // not a power of 2, denormal, or NaN
  1161  	}
  1162  	exp := b >> 52 & (1<<11 - 1)
  1163  	// exponent bias is 0x3ff.  So taking the reciprocal of a number
  1164  	// changes the exponent to 0x7fe-exp.
  1165  	switch exp {
  1166  	case 0:
  1167  		return false // ±0
  1168  	case 0x7ff:
  1169  		return false // ±inf
  1170  	case 0x7fe:
  1171  		return false // exponent is not representable
  1172  	default:
  1173  		return true
  1174  	}
  1175  }
  1176  
  1177  // reciprocalExact32 reports whether 1/c is exactly representable.
  1178  func reciprocalExact32(c float32) bool {
  1179  	b := math.Float32bits(c)
  1180  	man := b & (1<<23 - 1)
  1181  	if man != 0 {
  1182  		return false // not a power of 2, denormal, or NaN
  1183  	}
  1184  	exp := b >> 23 & (1<<8 - 1)
  1185  	// exponent bias is 0x7f.  So taking the reciprocal of a number
  1186  	// changes the exponent to 0xfe-exp.
  1187  	switch exp {
  1188  	case 0:
  1189  		return false // ±0
  1190  	case 0xff:
  1191  		return false // ±inf
  1192  	case 0xfe:
  1193  		return false // exponent is not representable
  1194  	default:
  1195  		return true
  1196  	}
  1197  }
  1198  
  1199  // check if an immediate can be directly encoded into an ARM's instruction
  1200  func isARMImmRot(v uint32) bool {
  1201  	for i := 0; i < 16; i++ {
  1202  		if v&^0xff == 0 {
  1203  			return true
  1204  		}
  1205  		v = v<<2 | v>>30
  1206  	}
  1207  
  1208  	return false
  1209  }
  1210  
  1211  // overlap reports whether the ranges given by the given offset and
  1212  // size pairs overlap.
  1213  func overlap(offset1, size1, offset2, size2 int64) bool {
  1214  	if offset1 >= offset2 && offset2+size2 > offset1 {
  1215  		return true
  1216  	}
  1217  	if offset2 >= offset1 && offset1+size1 > offset2 {
  1218  		return true
  1219  	}
  1220  	return false
  1221  }
  1222  
  1223  func areAdjacentOffsets(off1, off2, size int64) bool {
  1224  	return off1+size == off2 || off1 == off2+size
  1225  }
  1226  
  1227  // check if value zeroes out upper 32-bit of 64-bit register.
  1228  // depth limits recursion depth. In AMD64.rules 3 is used as limit,
  1229  // because it catches same amount of cases as 4.
  1230  func zeroUpper32Bits(x *Value, depth int) bool {
  1231  	switch x.Op {
  1232  	case OpAMD64MOVLconst, OpAMD64MOVLload, OpAMD64MOVLQZX, OpAMD64MOVLloadidx1,
  1233  		OpAMD64MOVWload, OpAMD64MOVWloadidx1, OpAMD64MOVBload, OpAMD64MOVBloadidx1,
  1234  		OpAMD64MOVLloadidx4, OpAMD64ADDLload, OpAMD64SUBLload, OpAMD64ANDLload,
  1235  		OpAMD64ORLload, OpAMD64XORLload, OpAMD64CVTTSD2SL,
  1236  		OpAMD64ADDL, OpAMD64ADDLconst, OpAMD64SUBL, OpAMD64SUBLconst,
  1237  		OpAMD64ANDL, OpAMD64ANDLconst, OpAMD64ORL, OpAMD64ORLconst,
  1238  		OpAMD64XORL, OpAMD64XORLconst, OpAMD64NEGL, OpAMD64NOTL,
  1239  		OpAMD64SHRL, OpAMD64SHRLconst, OpAMD64SARL, OpAMD64SARLconst,
  1240  		OpAMD64SHLL, OpAMD64SHLLconst:
  1241  		return true
  1242  	case OpArg:
  1243  		return x.Type.Width == 4
  1244  	case OpPhi, OpSelect0, OpSelect1:
  1245  		// Phis can use each-other as an arguments, instead of tracking visited values,
  1246  		// just limit recursion depth.
  1247  		if depth <= 0 {
  1248  			return false
  1249  		}
  1250  		for i := range x.Args {
  1251  			if !zeroUpper32Bits(x.Args[i], depth-1) {
  1252  				return false
  1253  			}
  1254  		}
  1255  		return true
  1256  
  1257  	}
  1258  	return false
  1259  }
  1260  
  1261  // zeroUpper48Bits is similar to zeroUpper32Bits, but for upper 48 bits
  1262  func zeroUpper48Bits(x *Value, depth int) bool {
  1263  	switch x.Op {
  1264  	case OpAMD64MOVWQZX, OpAMD64MOVWload, OpAMD64MOVWloadidx1, OpAMD64MOVWloadidx2:
  1265  		return true
  1266  	case OpArg:
  1267  		return x.Type.Width == 2
  1268  	case OpPhi, OpSelect0, OpSelect1:
  1269  		// Phis can use each-other as an arguments, instead of tracking visited values,
  1270  		// just limit recursion depth.
  1271  		if depth <= 0 {
  1272  			return false
  1273  		}
  1274  		for i := range x.Args {
  1275  			if !zeroUpper48Bits(x.Args[i], depth-1) {
  1276  				return false
  1277  			}
  1278  		}
  1279  		return true
  1280  
  1281  	}
  1282  	return false
  1283  }
  1284  
  1285  // zeroUpper56Bits is similar to zeroUpper32Bits, but for upper 56 bits
  1286  func zeroUpper56Bits(x *Value, depth int) bool {
  1287  	switch x.Op {
  1288  	case OpAMD64MOVBQZX, OpAMD64MOVBload, OpAMD64MOVBloadidx1:
  1289  		return true
  1290  	case OpArg:
  1291  		return x.Type.Width == 1
  1292  	case OpPhi, OpSelect0, OpSelect1:
  1293  		// Phis can use each-other as an arguments, instead of tracking visited values,
  1294  		// just limit recursion depth.
  1295  		if depth <= 0 {
  1296  			return false
  1297  		}
  1298  		for i := range x.Args {
  1299  			if !zeroUpper56Bits(x.Args[i], depth-1) {
  1300  				return false
  1301  			}
  1302  		}
  1303  		return true
  1304  
  1305  	}
  1306  	return false
  1307  }
  1308  
  1309  // isInlinableMemmove reports whether the given arch performs a Move of the given size
  1310  // faster than memmove. It will only return true if replacing the memmove with a Move is
  1311  // safe, either because Move is small or because the arguments are disjoint.
  1312  // This is used as a check for replacing memmove with Move ops.
  1313  func isInlinableMemmove(dst, src *Value, sz int64, c *Config) bool {
  1314  	// It is always safe to convert memmove into Move when its arguments are disjoint.
  1315  	// Move ops may or may not be faster for large sizes depending on how the platform
  1316  	// lowers them, so we only perform this optimization on platforms that we know to
  1317  	// have fast Move ops.
  1318  	switch c.arch {
  1319  	case "amd64":
  1320  		return sz <= 16 || (sz < 1024 && disjoint(dst, sz, src, sz))
  1321  	case "386", "arm64":
  1322  		return sz <= 8
  1323  	case "s390x", "ppc64", "ppc64le":
  1324  		return sz <= 8 || disjoint(dst, sz, src, sz)
  1325  	case "arm", "mips", "mips64", "mipsle", "mips64le":
  1326  		return sz <= 4
  1327  	}
  1328  	return false
  1329  }
  1330  
  1331  // logLargeCopy logs the occurrence of a large copy.
  1332  // The best place to do this is in the rewrite rules where the size of the move is easy to find.
  1333  // "Large" is arbitrarily chosen to be 128 bytes; this may change.
  1334  func logLargeCopy(v *Value, s int64) bool {
  1335  	if s < 128 {
  1336  		return true
  1337  	}
  1338  	if logopt.Enabled() {
  1339  		logopt.LogOpt(v.Pos, "copy", "lower", v.Block.Func.Name, fmt.Sprintf("%d bytes", s))
  1340  	}
  1341  	return true
  1342  }
  1343  
  1344  // hasSmallRotate reports whether the architecture has rotate instructions
  1345  // for sizes < 32-bit.  This is used to decide whether to promote some rotations.
  1346  func hasSmallRotate(c *Config) bool {
  1347  	switch c.arch {
  1348  	case "amd64", "386":
  1349  		return true
  1350  	default:
  1351  		return false
  1352  	}
  1353  }
  1354  
  1355  func newPPC64ShiftAuxInt(sh, mb, me, sz int64) int32 {
  1356  	if sh < 0 || sh >= sz {
  1357  		panic("PPC64 shift arg sh out of range")
  1358  	}
  1359  	if mb < 0 || mb >= sz {
  1360  		panic("PPC64 shift arg mb out of range")
  1361  	}
  1362  	if me < 0 || me >= sz {
  1363  		panic("PPC64 shift arg me out of range")
  1364  	}
  1365  	return int32(sh<<16 | mb<<8 | me)
  1366  }
  1367  
  1368  func GetPPC64Shiftsh(auxint int64) int64 {
  1369  	return int64(int8(auxint >> 16))
  1370  }
  1371  
  1372  func GetPPC64Shiftmb(auxint int64) int64 {
  1373  	return int64(int8(auxint >> 8))
  1374  }
  1375  
  1376  func GetPPC64Shiftme(auxint int64) int64 {
  1377  	return int64(int8(auxint))
  1378  }
  1379  
  1380  // Test if this value can encoded as a mask for a rlwinm like
  1381  // operation.  Masks can also extend from the msb and wrap to
  1382  // the lsb too.  That is, the valid masks are 32 bit strings
  1383  // of the form: 0..01..10..0 or 1..10..01..1 or 1...1
  1384  func isPPC64WordRotateMask(v64 int64) bool {
  1385  	// Isolate rightmost 1 (if none 0) and add.
  1386  	v := uint32(v64)
  1387  	vp := (v & -v) + v
  1388  	// Likewise, for the wrapping case.
  1389  	vn := ^v
  1390  	vpn := (vn & -vn) + vn
  1391  	return (v&vp == 0 || vn&vpn == 0) && v != 0
  1392  }
  1393  
  1394  // Compress mask and and shift into single value of the form
  1395  // me | mb<<8 | rotate<<16 | nbits<<24 where me and mb can
  1396  // be used to regenerate the input mask.
  1397  func encodePPC64RotateMask(rotate, mask, nbits int64) int64 {
  1398  	var mb, me, mbn, men int
  1399  
  1400  	// Determine boundaries and then decode them
  1401  	if mask == 0 || ^mask == 0 || rotate >= nbits {
  1402  		panic("Invalid PPC64 rotate mask")
  1403  	} else if nbits == 32 {
  1404  		mb = bits.LeadingZeros32(uint32(mask))
  1405  		me = 32 - bits.TrailingZeros32(uint32(mask))
  1406  		mbn = bits.LeadingZeros32(^uint32(mask))
  1407  		men = 32 - bits.TrailingZeros32(^uint32(mask))
  1408  	} else {
  1409  		mb = bits.LeadingZeros64(uint64(mask))
  1410  		me = 64 - bits.TrailingZeros64(uint64(mask))
  1411  		mbn = bits.LeadingZeros64(^uint64(mask))
  1412  		men = 64 - bits.TrailingZeros64(^uint64(mask))
  1413  	}
  1414  	// Check for a wrapping mask (e.g bits at 0 and 63)
  1415  	if mb == 0 && me == int(nbits) {
  1416  		// swap the inverted values
  1417  		mb, me = men, mbn
  1418  	}
  1419  
  1420  	return int64(me) | int64(mb<<8) | int64(rotate<<16) | int64(nbits<<24)
  1421  }
  1422  
  1423  // The inverse operation of encodePPC64RotateMask.  The values returned as
  1424  // mb and me satisfy the POWER ISA definition of MASK(x,y) where MASK(mb,me) = mask.
  1425  func DecodePPC64RotateMask(sauxint int64) (rotate, mb, me int64, mask uint64) {
  1426  	auxint := uint64(sauxint)
  1427  	rotate = int64((auxint >> 16) & 0xFF)
  1428  	mb = int64((auxint >> 8) & 0xFF)
  1429  	me = int64((auxint >> 0) & 0xFF)
  1430  	nbits := int64((auxint >> 24) & 0xFF)
  1431  	mask = ((1 << uint(nbits-mb)) - 1) ^ ((1 << uint(nbits-me)) - 1)
  1432  	if mb > me {
  1433  		mask = ^mask
  1434  	}
  1435  	if nbits == 32 {
  1436  		mask = uint64(uint32(mask))
  1437  	}
  1438  
  1439  	// Fixup ME to match ISA definition.  The second argument to MASK(..,me)
  1440  	// is inclusive.
  1441  	me = (me - 1) & (nbits - 1)
  1442  	return
  1443  }
  1444  
  1445  // This verifies that the mask is a set of
  1446  // consecutive bits including the least
  1447  // significant bit.
  1448  func isPPC64ValidShiftMask(v int64) bool {
  1449  	if (v != 0) && ((v+1)&v) == 0 {
  1450  		return true
  1451  	}
  1452  	return false
  1453  }
  1454  
  1455  func getPPC64ShiftMaskLength(v int64) int64 {
  1456  	return int64(bits.Len64(uint64(v)))
  1457  }
  1458  
  1459  // Decompose a shift right into an equivalent rotate/mask,
  1460  // and return mask & m.
  1461  func mergePPC64RShiftMask(m, s, nbits int64) int64 {
  1462  	smask := uint64((1<<uint(nbits))-1) >> uint(s)
  1463  	return m & int64(smask)
  1464  }
  1465  
  1466  // Combine (ANDconst [m] (SRWconst [s])) into (RLWINM [y]) or return 0
  1467  func mergePPC64AndSrwi(m, s int64) int64 {
  1468  	mask := mergePPC64RShiftMask(m, s, 32)
  1469  	if !isPPC64WordRotateMask(mask) {
  1470  		return 0
  1471  	}
  1472  	return encodePPC64RotateMask(32-s, mask, 32)
  1473  }
  1474  
  1475  // Test if a shift right feeding into a CLRLSLDI can be merged into RLWINM.
  1476  // Return the encoded RLWINM constant, or 0 if they cannot be merged.
  1477  func mergePPC64ClrlsldiSrw(sld, srw int64) int64 {
  1478  	mask_1 := uint64(0xFFFFFFFF >> uint(srw))
  1479  	// for CLRLSLDI, it's more convient to think of it as a mask left bits then rotate left.
  1480  	mask_2 := uint64(0xFFFFFFFFFFFFFFFF) >> uint(GetPPC64Shiftmb(int64(sld)))
  1481  
  1482  	// Rewrite mask to apply after the final left shift.
  1483  	mask_3 := (mask_1 & mask_2) << uint(GetPPC64Shiftsh(sld))
  1484  
  1485  	r_1 := 32 - srw
  1486  	r_2 := GetPPC64Shiftsh(sld)
  1487  	r_3 := (r_1 + r_2) & 31 // This can wrap.
  1488  
  1489  	if uint64(uint32(mask_3)) != mask_3 || mask_3 == 0 {
  1490  		return 0
  1491  	}
  1492  	return encodePPC64RotateMask(int64(r_3), int64(mask_3), 32)
  1493  }
  1494  
  1495  // Test if a RLWINM feeding into a CLRLSLDI can be merged into RLWINM.  Return
  1496  // the encoded RLWINM constant, or 0 if they cannot be merged.
  1497  func mergePPC64ClrlsldiRlwinm(sld int32, rlw int64) int64 {
  1498  	r_1, _, _, mask_1 := DecodePPC64RotateMask(rlw)
  1499  	// for CLRLSLDI, it's more convient to think of it as a mask left bits then rotate left.
  1500  	mask_2 := uint64(0xFFFFFFFFFFFFFFFF) >> uint(GetPPC64Shiftmb(int64(sld)))
  1501  
  1502  	// combine the masks, and adjust for the final left shift.
  1503  	mask_3 := (mask_1 & mask_2) << uint(GetPPC64Shiftsh(int64(sld)))
  1504  	r_2 := GetPPC64Shiftsh(int64(sld))
  1505  	r_3 := (r_1 + r_2) & 31 // This can wrap.
  1506  
  1507  	// Verify the result is still a valid bitmask of <= 32 bits.
  1508  	if !isPPC64WordRotateMask(int64(mask_3)) || uint64(uint32(mask_3)) != mask_3 {
  1509  		return 0
  1510  	}
  1511  	return encodePPC64RotateMask(r_3, int64(mask_3), 32)
  1512  }
  1513  
  1514  // Compute the encoded RLWINM constant from combining (SLDconst [sld] (SRWconst [srw] x)),
  1515  // or return 0 if they cannot be combined.
  1516  func mergePPC64SldiSrw(sld, srw int64) int64 {
  1517  	if sld > srw || srw >= 32 {
  1518  		return 0
  1519  	}
  1520  	mask_r := uint32(0xFFFFFFFF) >> uint(srw)
  1521  	mask_l := uint32(0xFFFFFFFF) >> uint(sld)
  1522  	mask := (mask_r & mask_l) << uint(sld)
  1523  	return encodePPC64RotateMask((32-srw+sld)&31, int64(mask), 32)
  1524  }
  1525  
  1526  // Convenience function to rotate a 32 bit constant value by another constant.
  1527  func rotateLeft32(v, rotate int64) int64 {
  1528  	return int64(bits.RotateLeft32(uint32(v), int(rotate)))
  1529  }
  1530  
  1531  // encodes the lsb and width for arm(64) bitfield ops into the expected auxInt format.
  1532  func armBFAuxInt(lsb, width int64) arm64BitField {
  1533  	if lsb < 0 || lsb > 63 {
  1534  		panic("ARM(64) bit field lsb constant out of range")
  1535  	}
  1536  	if width < 1 || width > 64 {
  1537  		panic("ARM(64) bit field width constant out of range")
  1538  	}
  1539  	return arm64BitField(width | lsb<<8)
  1540  }
  1541  
  1542  // returns the lsb part of the auxInt field of arm64 bitfield ops.
  1543  func (bfc arm64BitField) getARM64BFlsb() int64 {
  1544  	return int64(uint64(bfc) >> 8)
  1545  }
  1546  
  1547  // returns the width part of the auxInt field of arm64 bitfield ops.
  1548  func (bfc arm64BitField) getARM64BFwidth() int64 {
  1549  	return int64(bfc) & 0xff
  1550  }
  1551  
  1552  // checks if mask >> rshift applied at lsb is a valid arm64 bitfield op mask.
  1553  func isARM64BFMask(lsb, mask, rshift int64) bool {
  1554  	shiftedMask := int64(uint64(mask) >> uint64(rshift))
  1555  	return shiftedMask != 0 && isPowerOfTwo64(shiftedMask+1) && nto(shiftedMask)+lsb < 64
  1556  }
  1557  
  1558  // returns the bitfield width of mask >> rshift for arm64 bitfield ops
  1559  func arm64BFWidth(mask, rshift int64) int64 {
  1560  	shiftedMask := int64(uint64(mask) >> uint64(rshift))
  1561  	if shiftedMask == 0 {
  1562  		panic("ARM64 BF mask is zero")
  1563  	}
  1564  	return nto(shiftedMask)
  1565  }
  1566  
  1567  // sizeof returns the size of t in bytes.
  1568  // It will panic if t is not a *types.Type.
  1569  func sizeof(t interface{}) int64 {
  1570  	return t.(*types.Type).Size()
  1571  }
  1572  
  1573  // registerizable reports whether t is a primitive type that fits in
  1574  // a register. It assumes float64 values will always fit into registers
  1575  // even if that isn't strictly true.
  1576  func registerizable(b *Block, typ *types.Type) bool {
  1577  	if typ.IsPtrShaped() || typ.IsFloat() {
  1578  		return true
  1579  	}
  1580  	if typ.IsInteger() {
  1581  		return typ.Size() <= b.Func.Config.RegSize
  1582  	}
  1583  	return false
  1584  }
  1585  
  1586  // needRaceCleanup reports whether this call to racefuncenter/exit isn't needed.
  1587  func needRaceCleanup(sym *AuxCall, v *Value) bool {
  1588  	f := v.Block.Func
  1589  	if !f.Config.Race {
  1590  		return false
  1591  	}
  1592  	if !isSameCall(sym, "runtime.racefuncenter") && !isSameCall(sym, "runtime.racefuncenterfp") && !isSameCall(sym, "runtime.racefuncexit") {
  1593  		return false
  1594  	}
  1595  	for _, b := range f.Blocks {
  1596  		for _, v := range b.Values {
  1597  			switch v.Op {
  1598  			case OpStaticCall:
  1599  				// Check for racefuncenter/racefuncenterfp will encounter racefuncexit and vice versa.
  1600  				// Allow calls to panic*
  1601  				s := v.Aux.(*AuxCall).Fn.String()
  1602  				switch s {
  1603  				case "runtime.racefuncenter", "runtime.racefuncenterfp", "runtime.racefuncexit",
  1604  					"runtime.panicdivide", "runtime.panicwrap",
  1605  					"runtime.panicshift":
  1606  					continue
  1607  				}
  1608  				// If we encountered any call, we need to keep racefunc*,
  1609  				// for accurate stacktraces.
  1610  				return false
  1611  			case OpPanicBounds, OpPanicExtend:
  1612  				// Note: these are panic generators that are ok (like the static calls above).
  1613  			case OpClosureCall, OpInterCall:
  1614  				// We must keep the race functions if there are any other call types.
  1615  				return false
  1616  			}
  1617  		}
  1618  	}
  1619  	if isSameCall(sym, "runtime.racefuncenter") {
  1620  		// If we're removing racefuncenter, remove its argument as well.
  1621  		if v.Args[0].Op != OpStore {
  1622  			return false
  1623  		}
  1624  		mem := v.Args[0].Args[2]
  1625  		v.Args[0].reset(OpCopy)
  1626  		v.Args[0].AddArg(mem)
  1627  	}
  1628  	return true
  1629  }
  1630  
  1631  // symIsRO reports whether sym is a read-only global.
  1632  func symIsRO(sym interface{}) bool {
  1633  	lsym := sym.(*obj.LSym)
  1634  	return lsym.Type == objabi.SRODATA && len(lsym.R) == 0
  1635  }
  1636  
  1637  // symIsROZero reports whether sym is a read-only global whose data contains all zeros.
  1638  func symIsROZero(sym Sym) bool {
  1639  	lsym := sym.(*obj.LSym)
  1640  	if lsym.Type != objabi.SRODATA || len(lsym.R) != 0 {
  1641  		return false
  1642  	}
  1643  	for _, b := range lsym.P {
  1644  		if b != 0 {
  1645  			return false
  1646  		}
  1647  	}
  1648  	return true
  1649  }
  1650  
  1651  // read8 reads one byte from the read-only global sym at offset off.
  1652  func read8(sym interface{}, off int64) uint8 {
  1653  	lsym := sym.(*obj.LSym)
  1654  	if off >= int64(len(lsym.P)) || off < 0 {
  1655  		// Invalid index into the global sym.
  1656  		// This can happen in dead code, so we don't want to panic.
  1657  		// Just return any value, it will eventually get ignored.
  1658  		// See issue 29215.
  1659  		return 0
  1660  	}
  1661  	return lsym.P[off]
  1662  }
  1663  
  1664  // read16 reads two bytes from the read-only global sym at offset off.
  1665  func read16(sym interface{}, off int64, byteorder binary.ByteOrder) uint16 {
  1666  	lsym := sym.(*obj.LSym)
  1667  	// lsym.P is written lazily.
  1668  	// Bytes requested after the end of lsym.P are 0.
  1669  	var src []byte
  1670  	if 0 <= off && off < int64(len(lsym.P)) {
  1671  		src = lsym.P[off:]
  1672  	}
  1673  	buf := make([]byte, 2)
  1674  	copy(buf, src)
  1675  	return byteorder.Uint16(buf)
  1676  }
  1677  
  1678  // read32 reads four bytes from the read-only global sym at offset off.
  1679  func read32(sym interface{}, off int64, byteorder binary.ByteOrder) uint32 {
  1680  	lsym := sym.(*obj.LSym)
  1681  	var src []byte
  1682  	if 0 <= off && off < int64(len(lsym.P)) {
  1683  		src = lsym.P[off:]
  1684  	}
  1685  	buf := make([]byte, 4)
  1686  	copy(buf, src)
  1687  	return byteorder.Uint32(buf)
  1688  }
  1689  
  1690  // read64 reads eight bytes from the read-only global sym at offset off.
  1691  func read64(sym interface{}, off int64, byteorder binary.ByteOrder) uint64 {
  1692  	lsym := sym.(*obj.LSym)
  1693  	var src []byte
  1694  	if 0 <= off && off < int64(len(lsym.P)) {
  1695  		src = lsym.P[off:]
  1696  	}
  1697  	buf := make([]byte, 8)
  1698  	copy(buf, src)
  1699  	return byteorder.Uint64(buf)
  1700  }
  1701  
  1702  // sequentialAddresses reports true if it can prove that x + n == y
  1703  func sequentialAddresses(x, y *Value, n int64) bool {
  1704  	if x.Op == Op386ADDL && y.Op == Op386LEAL1 && y.AuxInt == n && y.Aux == nil &&
  1705  		(x.Args[0] == y.Args[0] && x.Args[1] == y.Args[1] ||
  1706  			x.Args[0] == y.Args[1] && x.Args[1] == y.Args[0]) {
  1707  		return true
  1708  	}
  1709  	if x.Op == Op386LEAL1 && y.Op == Op386LEAL1 && y.AuxInt == x.AuxInt+n && x.Aux == y.Aux &&
  1710  		(x.Args[0] == y.Args[0] && x.Args[1] == y.Args[1] ||
  1711  			x.Args[0] == y.Args[1] && x.Args[1] == y.Args[0]) {
  1712  		return true
  1713  	}
  1714  	if x.Op == OpAMD64ADDQ && y.Op == OpAMD64LEAQ1 && y.AuxInt == n && y.Aux == nil &&
  1715  		(x.Args[0] == y.Args[0] && x.Args[1] == y.Args[1] ||
  1716  			x.Args[0] == y.Args[1] && x.Args[1] == y.Args[0]) {
  1717  		return true
  1718  	}
  1719  	if x.Op == OpAMD64LEAQ1 && y.Op == OpAMD64LEAQ1 && y.AuxInt == x.AuxInt+n && x.Aux == y.Aux &&
  1720  		(x.Args[0] == y.Args[0] && x.Args[1] == y.Args[1] ||
  1721  			x.Args[0] == y.Args[1] && x.Args[1] == y.Args[0]) {
  1722  		return true
  1723  	}
  1724  	return false
  1725  }
  1726  
  1727  // flagConstant represents the result of a compile-time comparison.
  1728  // The sense of these flags does not necessarily represent the hardware's notion
  1729  // of a flags register - these are just a compile-time construct.
  1730  // We happen to match the semantics to those of arm/arm64.
  1731  // Note that these semantics differ from x86: the carry flag has the opposite
  1732  // sense on a subtraction!
  1733  //   On amd64, C=1 represents a borrow, e.g. SBB on amd64 does x - y - C.
  1734  //   On arm64, C=0 represents a borrow, e.g. SBC on arm64 does x - y - ^C.
  1735  //    (because it does x + ^y + C).
  1736  // See https://en.wikipedia.org/wiki/Carry_flag#Vs._borrow_flag
  1737  type flagConstant uint8
  1738  
  1739  // N reports whether the result of an operation is negative (high bit set).
  1740  func (fc flagConstant) N() bool {
  1741  	return fc&1 != 0
  1742  }
  1743  
  1744  // Z reports whether the result of an operation is 0.
  1745  func (fc flagConstant) Z() bool {
  1746  	return fc&2 != 0
  1747  }
  1748  
  1749  // C reports whether an unsigned add overflowed (carry), or an
  1750  // unsigned subtract did not underflow (borrow).
  1751  func (fc flagConstant) C() bool {
  1752  	return fc&4 != 0
  1753  }
  1754  
  1755  // V reports whether a signed operation overflowed or underflowed.
  1756  func (fc flagConstant) V() bool {
  1757  	return fc&8 != 0
  1758  }
  1759  
  1760  func (fc flagConstant) eq() bool {
  1761  	return fc.Z()
  1762  }
  1763  func (fc flagConstant) ne() bool {
  1764  	return !fc.Z()
  1765  }
  1766  func (fc flagConstant) lt() bool {
  1767  	return fc.N() != fc.V()
  1768  }
  1769  func (fc flagConstant) le() bool {
  1770  	return fc.Z() || fc.lt()
  1771  }
  1772  func (fc flagConstant) gt() bool {
  1773  	return !fc.Z() && fc.ge()
  1774  }
  1775  func (fc flagConstant) ge() bool {
  1776  	return fc.N() == fc.V()
  1777  }
  1778  func (fc flagConstant) ult() bool {
  1779  	return !fc.C()
  1780  }
  1781  func (fc flagConstant) ule() bool {
  1782  	return fc.Z() || fc.ult()
  1783  }
  1784  func (fc flagConstant) ugt() bool {
  1785  	return !fc.Z() && fc.uge()
  1786  }
  1787  func (fc flagConstant) uge() bool {
  1788  	return fc.C()
  1789  }
  1790  
  1791  func (fc flagConstant) ltNoov() bool {
  1792  	return fc.lt() && !fc.V()
  1793  }
  1794  func (fc flagConstant) leNoov() bool {
  1795  	return fc.le() && !fc.V()
  1796  }
  1797  func (fc flagConstant) gtNoov() bool {
  1798  	return fc.gt() && !fc.V()
  1799  }
  1800  func (fc flagConstant) geNoov() bool {
  1801  	return fc.ge() && !fc.V()
  1802  }
  1803  
  1804  func (fc flagConstant) String() string {
  1805  	return fmt.Sprintf("N=%v,Z=%v,C=%v,V=%v", fc.N(), fc.Z(), fc.C(), fc.V())
  1806  }
  1807  
  1808  type flagConstantBuilder struct {
  1809  	N bool
  1810  	Z bool
  1811  	C bool
  1812  	V bool
  1813  }
  1814  
  1815  func (fcs flagConstantBuilder) encode() flagConstant {
  1816  	var fc flagConstant
  1817  	if fcs.N {
  1818  		fc |= 1
  1819  	}
  1820  	if fcs.Z {
  1821  		fc |= 2
  1822  	}
  1823  	if fcs.C {
  1824  		fc |= 4
  1825  	}
  1826  	if fcs.V {
  1827  		fc |= 8
  1828  	}
  1829  	return fc
  1830  }
  1831  
  1832  // Note: addFlags(x,y) != subFlags(x,-y) in some situations:
  1833  //  - the results of the C flag are different
  1834  //  - the results of the V flag when y==minint are different
  1835  
  1836  // addFlags64 returns the flags that would be set from computing x+y.
  1837  func addFlags64(x, y int64) flagConstant {
  1838  	var fcb flagConstantBuilder
  1839  	fcb.Z = x+y == 0
  1840  	fcb.N = x+y < 0
  1841  	fcb.C = uint64(x+y) < uint64(x)
  1842  	fcb.V = x >= 0 && y >= 0 && x+y < 0 || x < 0 && y < 0 && x+y >= 0
  1843  	return fcb.encode()
  1844  }
  1845  
  1846  // subFlags64 returns the flags that would be set from computing x-y.
  1847  func subFlags64(x, y int64) flagConstant {
  1848  	var fcb flagConstantBuilder
  1849  	fcb.Z = x-y == 0
  1850  	fcb.N = x-y < 0
  1851  	fcb.C = uint64(y) <= uint64(x) // This code follows the arm carry flag model.
  1852  	fcb.V = x >= 0 && y < 0 && x-y < 0 || x < 0 && y >= 0 && x-y >= 0
  1853  	return fcb.encode()
  1854  }
  1855  
  1856  // addFlags32 returns the flags that would be set from computing x+y.
  1857  func addFlags32(x, y int32) flagConstant {
  1858  	var fcb flagConstantBuilder
  1859  	fcb.Z = x+y == 0
  1860  	fcb.N = x+y < 0
  1861  	fcb.C = uint32(x+y) < uint32(x)
  1862  	fcb.V = x >= 0 && y >= 0 && x+y < 0 || x < 0 && y < 0 && x+y >= 0
  1863  	return fcb.encode()
  1864  }
  1865  
  1866  // subFlags32 returns the flags that would be set from computing x-y.
  1867  func subFlags32(x, y int32) flagConstant {
  1868  	var fcb flagConstantBuilder
  1869  	fcb.Z = x-y == 0
  1870  	fcb.N = x-y < 0
  1871  	fcb.C = uint32(y) <= uint32(x) // This code follows the arm carry flag model.
  1872  	fcb.V = x >= 0 && y < 0 && x-y < 0 || x < 0 && y >= 0 && x-y >= 0
  1873  	return fcb.encode()
  1874  }
  1875  
  1876  // logicFlags64 returns flags set to the sign/zeroness of x.
  1877  // C and V are set to false.
  1878  func logicFlags64(x int64) flagConstant {
  1879  	var fcb flagConstantBuilder
  1880  	fcb.Z = x == 0
  1881  	fcb.N = x < 0
  1882  	return fcb.encode()
  1883  }
  1884  
  1885  // logicFlags32 returns flags set to the sign/zeroness of x.
  1886  // C and V are set to false.
  1887  func logicFlags32(x int32) flagConstant {
  1888  	var fcb flagConstantBuilder
  1889  	fcb.Z = x == 0
  1890  	fcb.N = x < 0
  1891  	return fcb.encode()
  1892  }
  1893  

View as plain text