config.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/abi"
     9  	"cmd/compile/internal/base"
    10  	"cmd/compile/internal/ir"
    11  	"cmd/compile/internal/ssa/ssabase"
    12  	"cmd/compile/internal/ssa/ssaop"
    13  	"cmd/compile/internal/types"
    14  	"cmd/internal/obj"
    15  	"cmd/internal/src"
    16  )
    17  
    18  type (
    19  	BlockRewriter func(*Block) bool
    20  	ValueRewriter func(*Value) bool
    21  )
    22  
    23  // A Config holds readonly compilation information.
    24  // It is created once, early during compilation,
    25  // and shared across all compilations.
    26  type Config struct {
    27  	Arch           string // "amd64", etc.
    28  	PtrSize        int64  // 4 or 8; copy of cmd/internal/sys.Arch.PtrSize
    29  	RegSize        int64  // 4 or 8; copy of cmd/internal/sys.Arch.RegSize
    30  	Types          Types
    31  	LowerBlock     BlockRewriter      // block lowering function, first round
    32  	LowerValue     ValueRewriter      // value lowering function, first round
    33  	LateLowerBlock BlockRewriter      // block lowering function that needs to be run after the first round; only used on some architectures
    34  	LateLowerValue ValueRewriter      // value lowering function that needs to be run after the first round; only used on some architectures
    35  	SplitLoad      ValueRewriter      // function for splitting merged load ops; only used on some architectures
    36  	Registers      []ssabase.Register // machine registers
    37  	GpRegMask      ssaop.RegMask      // general purpose integer register mask
    38  	FpRegMask      ssaop.RegMask      // floating point register mask
    39  	Fp32RegMask    ssaop.RegMask      // floating point register mask
    40  	Fp64RegMask    ssaop.RegMask      // floating point register mask
    41  	SimdRegMask    ssaop.RegMask      // simd register mask; may be same as fpRegMask
    42  	SpecialRegMask ssaop.RegMask      // special register mask
    43  	IntParamRegs   []int8             // register numbers of integer param (in/out) registers
    44  	FloatParamRegs []int8             // register numbers of floating param (in/out) registers
    45  	ABI1           *abi.ABIConfig     // "ABIInternal" under development // TODO change comment when this becomes current
    46  	ABI0           *abi.ABIConfig
    47  	FPReg          int8      // register number of frame pointer, -1 if not used
    48  	LinkReg        int8      // register number of link register if it is a general purpose register, -1 if not used
    49  	HasGReg        bool      // has hardware g register
    50  	Ctxt           *obj.Link // Generic arch information
    51  	Optimize       bool      // Do optimization
    52  	SoftFloat      bool      //
    53  	Race           bool      // race detector enabled
    54  	BigEndian      bool      //
    55  	UnalignedOK    bool      // Unaligned loads/stores are ok
    56  	HaveBswap64    bool      // architecture implements Bswap64
    57  	HaveBswap32    bool      // architecture implements Bswap32
    58  	HaveBswap16    bool      // architecture implements Bswap16
    59  	HaveCondSelect bool      // architecture implements CondSelect
    60  
    61  	// MulRecipes[x] = function to build v * x from v.
    62  	MulRecipes map[int64]mulRecipe
    63  }
    64  
    65  type Frontend interface {
    66  	Logger
    67  
    68  	// StringData returns a symbol pointing to the given string's contents.
    69  	StringData(string) *obj.LSym
    70  
    71  	// Given the name for a compound type, returns the name we should use
    72  	// for the parts of that compound type.
    73  	SplitSlot(parent *LocalSlot, suffix string, offset int64, t *types.Type) LocalSlot
    74  
    75  	// Syslook returns a symbol of the runtime function/variable with the
    76  	// given name.
    77  	Syslook(string) *obj.LSym
    78  
    79  	// UseWriteBarrier reports whether write barrier is enabled
    80  	UseWriteBarrier() bool
    81  
    82  	// Func returns the ir.Func of the function being compiled.
    83  	Func() *ir.Func
    84  }
    85  
    86  type Logger interface {
    87  	// Logf logs a message from the compiler.
    88  	Logf(string, ...any)
    89  
    90  	// Log reports whether logging is not a no-op
    91  	// some logging calls account for more than a few heap allocations.
    92  	Log() bool
    93  
    94  	// Fatalf reports a compiler error and exits.
    95  	Fatalf(pos src.XPos, msg string, args ...any)
    96  
    97  	// Warnl writes compiler messages in the form expected by "errorcheck" tests
    98  	Warnl(pos src.XPos, fmt_ string, args ...any)
    99  
   100  	// Forwards the Debug flags from gc
   101  	Debug_checknil() bool
   102  }
   103  
   104  type Types struct {
   105  	Bool       *types.Type
   106  	Int8       *types.Type
   107  	Int16      *types.Type
   108  	Int32      *types.Type
   109  	Int64      *types.Type
   110  	UInt8      *types.Type
   111  	UInt16     *types.Type
   112  	UInt32     *types.Type
   113  	UInt64     *types.Type
   114  	Int        *types.Type
   115  	Float32    *types.Type
   116  	Float64    *types.Type
   117  	UInt       *types.Type
   118  	Uintptr    *types.Type
   119  	String     *types.Type
   120  	BytePtr    *types.Type // TODO: use unsafe.Pointer instead?
   121  	Int32Ptr   *types.Type
   122  	UInt32Ptr  *types.Type
   123  	IntPtr     *types.Type
   124  	UintptrPtr *types.Type
   125  	Float32Ptr *types.Type
   126  	Float64Ptr *types.Type
   127  	BytePtrPtr *types.Type
   128  	Vec128     *types.Type
   129  	Vec256     *types.Type
   130  	Vec512     *types.Type
   131  	Mask       *types.Type
   132  }
   133  
   134  type mulRecipe struct {
   135  	cost  int
   136  	Build func(*Value, *Value) *Value // build(m, v) returns v * x built at m.
   137  }
   138  
   139  // NewTypes creates and populates a Types.
   140  func NewTypes() *Types {
   141  	t := new(Types)
   142  	t.SetTypPtrs()
   143  	return t
   144  }
   145  
   146  // SetTypPtrs populates t.
   147  func (t *Types) SetTypPtrs() {
   148  	t.Bool = types.Types[types.TBOOL]
   149  	t.Int8 = types.Types[types.TINT8]
   150  	t.Int16 = types.Types[types.TINT16]
   151  	t.Int32 = types.Types[types.TINT32]
   152  	t.Int64 = types.Types[types.TINT64]
   153  	t.UInt8 = types.Types[types.TUINT8]
   154  	t.UInt16 = types.Types[types.TUINT16]
   155  	t.UInt32 = types.Types[types.TUINT32]
   156  	t.UInt64 = types.Types[types.TUINT64]
   157  	t.Int = types.Types[types.TINT]
   158  	t.Float32 = types.Types[types.TFLOAT32]
   159  	t.Float64 = types.Types[types.TFLOAT64]
   160  	t.UInt = types.Types[types.TUINT]
   161  	t.Uintptr = types.Types[types.TUINTPTR]
   162  	t.String = types.Types[types.TSTRING]
   163  	t.BytePtr = types.NewPtr(types.Types[types.TUINT8])
   164  	t.Int32Ptr = types.NewPtr(types.Types[types.TINT32])
   165  	t.UInt32Ptr = types.NewPtr(types.Types[types.TUINT32])
   166  	t.IntPtr = types.NewPtr(types.Types[types.TINT])
   167  	t.UintptrPtr = types.NewPtr(types.Types[types.TUINTPTR])
   168  	t.Float32Ptr = types.NewPtr(types.Types[types.TFLOAT32])
   169  	t.Float64Ptr = types.NewPtr(types.Types[types.TFLOAT64])
   170  	t.BytePtrPtr = types.NewPtr(types.NewPtr(types.Types[types.TUINT8]))
   171  	t.Vec128 = types.TypeVec128
   172  	t.Vec256 = types.TypeVec256
   173  	t.Vec512 = types.TypeVec512
   174  	t.Mask = types.TypeMask
   175  }
   176  
   177  func (c *Config) HaveByteSwap(size int64) bool {
   178  	switch size {
   179  	case 8:
   180  		return c.HaveBswap64
   181  	case 4:
   182  		return c.HaveBswap32
   183  	case 2:
   184  		return c.HaveBswap16
   185  	default:
   186  		base.Fatalf("bad size %d\n", size)
   187  		return false
   188  	}
   189  }
   190  
   191  func (c *Config) BuildRecipes(arch string) {
   192  	// Information for strength-reducing multiplies.
   193  	type linearCombo struct {
   194  		// we can compute a*x+b*y in one instruction
   195  		a, b int64
   196  		// cost, in arbitrary units (tenths of cycles, usually)
   197  		cost int
   198  		// builds SSA value for a*x+b*y. Use the position
   199  		// information from m.
   200  		build func(m, x, y *Value) *Value
   201  	}
   202  
   203  	// List all the linear combination instructions we have.
   204  	var linearCombos []linearCombo
   205  	r := func(a, b int64, cost int, build func(m, x, y *Value) *Value) {
   206  		linearCombos = append(linearCombos, linearCombo{a: a, b: b, cost: cost, build: build})
   207  	}
   208  	var mulCost int
   209  	switch arch {
   210  	case "amd64":
   211  		// Assumes that the following costs from https://gmplib.org/~tege/x86-timing.pdf:
   212  		//    1 - addq, shlq, leaq, negq, subq
   213  		//    3 - imulq
   214  		// These costs limit the rewrites to two instructions.
   215  		// Operations which have to happen in place (and thus
   216  		// may require a reg-reg move) score slightly higher.
   217  		mulCost = 30
   218  		// add
   219  		r(1, 1, 10,
   220  			func(m, x, y *Value) *Value {
   221  				v := m.Block.NewValue2(m.Pos, ssaop.OpAMD64ADDQ, m.Type, x, y)
   222  				if m.Type.Size() == 4 {
   223  					v.Op = ssaop.OpAMD64ADDL
   224  				}
   225  				return v
   226  			})
   227  		// neg
   228  		r(-1, 0, 11,
   229  			func(m, x, y *Value) *Value {
   230  				v := m.Block.NewValue1(m.Pos, ssaop.OpAMD64NEGQ, m.Type, x)
   231  				if m.Type.Size() == 4 {
   232  					v.Op = ssaop.OpAMD64NEGL
   233  				}
   234  				return v
   235  			})
   236  		// sub
   237  		r(1, -1, 11,
   238  			func(m, x, y *Value) *Value {
   239  				v := m.Block.NewValue2(m.Pos, ssaop.OpAMD64SUBQ, m.Type, x, y)
   240  				if m.Type.Size() == 4 {
   241  					v.Op = ssaop.OpAMD64SUBL
   242  				}
   243  				return v
   244  			})
   245  		// lea
   246  		r(1, 2, 10,
   247  			func(m, x, y *Value) *Value {
   248  				v := m.Block.NewValue2(m.Pos, ssaop.OpAMD64LEAQ2, m.Type, x, y)
   249  				if m.Type.Size() == 4 {
   250  					v.Op = ssaop.OpAMD64LEAL2
   251  				}
   252  				return v
   253  			})
   254  		r(1, 4, 10,
   255  			func(m, x, y *Value) *Value {
   256  				v := m.Block.NewValue2(m.Pos, ssaop.OpAMD64LEAQ4, m.Type, x, y)
   257  				if m.Type.Size() == 4 {
   258  					v.Op = ssaop.OpAMD64LEAL4
   259  				}
   260  				return v
   261  			})
   262  		r(1, 8, 10,
   263  			func(m, x, y *Value) *Value {
   264  				v := m.Block.NewValue2(m.Pos, ssaop.OpAMD64LEAQ8, m.Type, x, y)
   265  				if m.Type.Size() == 4 {
   266  					v.Op = ssaop.OpAMD64LEAL8
   267  				}
   268  				return v
   269  			})
   270  		// regular shifts
   271  		for i := 2; i < 64; i++ {
   272  			r(1<<i, 0, 11,
   273  				func(m, x, y *Value) *Value {
   274  					v := m.Block.NewValue1I(m.Pos, ssaop.OpAMD64SHLQconst, m.Type, int64(i), x)
   275  					if m.Type.Size() == 4 {
   276  						v.Op = ssaop.OpAMD64SHLLconst
   277  					}
   278  					return v
   279  				})
   280  		}
   281  
   282  	case "arm64":
   283  		// Rationale (for M2 ultra):
   284  		// - multiply is 3 cycles.
   285  		// - add/neg/sub/shift are 1 cycle.
   286  		// - add/neg/sub+shiftLL are 2 cycles.
   287  		// We break ties against the multiply because using a
   288  		// multiply also needs to load the constant into a register.
   289  		// (It's 3 cycles and 2 instructions either way, but the
   290  		// linear combo one might use 1 less register.)
   291  		// The multiply constant might get lifted out of a loop though. Hmm....
   292  		// Other arm64 chips have different tradeoffs.
   293  		// Some chip's add+shift instructions are 1 cycle for shifts up to 4
   294  		// and 2 cycles for shifts bigger than 4. So weight the larger shifts
   295  		// a bit more.
   296  		// TODO: figure out a happy medium.
   297  		mulCost = 35
   298  		// add
   299  		r(1, 1, 10,
   300  			func(m, x, y *Value) *Value {
   301  				return m.Block.NewValue2(m.Pos, ssaop.OpARM64ADD, m.Type, x, y)
   302  			})
   303  		// neg
   304  		r(-1, 0, 10,
   305  			func(m, x, y *Value) *Value {
   306  				return m.Block.NewValue1(m.Pos, ssaop.OpARM64NEG, m.Type, x)
   307  			})
   308  		// sub
   309  		r(1, -1, 10,
   310  			func(m, x, y *Value) *Value {
   311  				return m.Block.NewValue2(m.Pos, ssaop.OpARM64SUB, m.Type, x, y)
   312  			})
   313  		// ADDshiftLL
   314  		for i := 1; i < 64; i++ {
   315  			c := 20
   316  			if i > 4 {
   317  				c++
   318  			}
   319  			r(1, 1<<i, c,
   320  				func(m, x, y *Value) *Value {
   321  					return m.Block.NewValue2I(m.Pos, ssaop.OpARM64ADDshiftLL, m.Type, int64(i), x, y)
   322  				})
   323  		}
   324  		// NEGshiftLL
   325  		for i := 1; i < 64; i++ {
   326  			c := 20
   327  			if i > 4 {
   328  				c++
   329  			}
   330  			r(-1<<i, 0, c,
   331  				func(m, x, y *Value) *Value {
   332  					return m.Block.NewValue1I(m.Pos, ssaop.OpARM64NEGshiftLL, m.Type, int64(i), x)
   333  				})
   334  		}
   335  		// SUBshiftLL
   336  		for i := 1; i < 64; i++ {
   337  			c := 20
   338  			if i > 4 {
   339  				c++
   340  			}
   341  			r(1, -1<<i, c,
   342  				func(m, x, y *Value) *Value {
   343  					return m.Block.NewValue2I(m.Pos, ssaop.OpARM64SUBshiftLL, m.Type, int64(i), x, y)
   344  				})
   345  		}
   346  		// regular shifts
   347  		for i := 1; i < 64; i++ {
   348  			c := 10
   349  			if i == 1 {
   350  				// Prefer x<<1 over x+x.
   351  				// Note that we eventually reverse this decision in ARM64latelower.rules,
   352  				// but this makes shift combining rules in ARM64.rules simpler.
   353  				c--
   354  			}
   355  			r(1<<i, 0, c,
   356  				func(m, x, y *Value) *Value {
   357  					return m.Block.NewValue1I(m.Pos, ssaop.OpARM64SLLconst, m.Type, int64(i), x)
   358  				})
   359  		}
   360  	case "loong64":
   361  		// - multiply is 4 cycles.
   362  		// - add/sub/shift/alsl are 1 cycle.
   363  		// On loong64, using a multiply also needs to load the constant into a register.
   364  		// TODO: figure out a happy medium.
   365  		mulCost = 45
   366  
   367  		// add
   368  		r(1, 1, 10,
   369  			func(m, x, y *Value) *Value {
   370  				return m.Block.NewValue2(m.Pos, ssaop.OpLOONG64ADDV, m.Type, x, y)
   371  			})
   372  		// neg
   373  		r(-1, 0, 10,
   374  			func(m, x, y *Value) *Value {
   375  				return m.Block.NewValue1(m.Pos, ssaop.OpLOONG64NEGV, m.Type, x)
   376  			})
   377  		// sub
   378  		r(1, -1, 10,
   379  			func(m, x, y *Value) *Value {
   380  				return m.Block.NewValue2(m.Pos, ssaop.OpLOONG64SUBV, m.Type, x, y)
   381  			})
   382  
   383  		// ADDshiftLLV
   384  		for i := 1; i < 5; i++ {
   385  			c := 10
   386  			r(1, 1<<i, c,
   387  				func(m, x, y *Value) *Value {
   388  					return m.Block.NewValue2I(m.Pos, ssaop.OpLOONG64ADDshiftLLV, m.Type, int64(i), x, y)
   389  				})
   390  		}
   391  
   392  		// regular shifts
   393  		for i := 1; i < 64; i++ {
   394  			c := 10
   395  			if i == 1 {
   396  				// Prefer x<<1 over x+x.
   397  				// Note that we eventually reverse this decision in LOONG64latelower.rules,
   398  				// but this makes shift combining rules in LOONG64.rules simpler.
   399  				c--
   400  			}
   401  			r(1<<i, 0, c,
   402  				func(m, x, y *Value) *Value {
   403  					return m.Block.NewValue1I(m.Pos, ssaop.OpLOONG64SLLVconst, m.Type, int64(i), x)
   404  				})
   405  		}
   406  	}
   407  
   408  	c.MulRecipes = map[int64]mulRecipe{}
   409  
   410  	// Single-instruction recipes.
   411  	// The only option for the input value(s) is v.
   412  	for _, combo := range linearCombos {
   413  		x := combo.a + combo.b
   414  		cost := combo.cost
   415  		old := c.MulRecipes[x]
   416  		if (old.Build == nil || cost < old.cost) && cost < mulCost {
   417  			c.MulRecipes[x] = mulRecipe{cost: cost, Build: func(m, v *Value) *Value {
   418  				return combo.build(m, v, v)
   419  			}}
   420  		}
   421  	}
   422  	// Two-instruction recipes.
   423  	// A: Both of the outer's inputs are from the same single-instruction recipe.
   424  	// B: First input is v and the second is from a single-instruction recipe.
   425  	// C: Second input is v and the first is from a single-instruction recipe.
   426  	// A is slightly preferred because it often needs 1 less register, so it
   427  	// goes first.
   428  
   429  	// A
   430  	for _, inner := range linearCombos {
   431  		for _, outer := range linearCombos {
   432  			x := (inner.a + inner.b) * (outer.a + outer.b)
   433  			cost := inner.cost + outer.cost
   434  			old := c.MulRecipes[x]
   435  			if (old.Build == nil || cost < old.cost) && cost < mulCost {
   436  				c.MulRecipes[x] = mulRecipe{cost: cost, Build: func(m, v *Value) *Value {
   437  					v = inner.build(m, v, v)
   438  					return outer.build(m, v, v)
   439  				}}
   440  			}
   441  		}
   442  	}
   443  
   444  	// B
   445  	for _, inner := range linearCombos {
   446  		for _, outer := range linearCombos {
   447  			x := outer.a + outer.b*(inner.a+inner.b)
   448  			cost := inner.cost + outer.cost
   449  			old := c.MulRecipes[x]
   450  			if (old.Build == nil || cost < old.cost) && cost < mulCost {
   451  				c.MulRecipes[x] = mulRecipe{cost: cost, Build: func(m, v *Value) *Value {
   452  					return outer.build(m, v, inner.build(m, v, v))
   453  				}}
   454  			}
   455  		}
   456  	}
   457  
   458  	// C
   459  	for _, inner := range linearCombos {
   460  		for _, outer := range linearCombos {
   461  			x := outer.a*(inner.a+inner.b) + outer.b
   462  			cost := inner.cost + outer.cost
   463  			old := c.MulRecipes[x]
   464  			if (old.Build == nil || cost < old.cost) && cost < mulCost {
   465  				c.MulRecipes[x] = mulRecipe{cost: cost, Build: func(m, v *Value) *Value {
   466  					return outer.build(m, inner.build(m, v, v), v)
   467  				}}
   468  			}
   469  		}
   470  	}
   471  
   472  	// Currently we only process 3 linear combination instructions for loong64.
   473  	if arch == "loong64" {
   474  		// Three-instruction recipes.
   475  		// D: The first and the second are all single-instruction recipes, and they are also the third's inputs.
   476  		// E: The first single-instruction is the second's input, and the second is the third's input.
   477  
   478  		// D
   479  		for _, first := range linearCombos {
   480  			for _, second := range linearCombos {
   481  				for _, third := range linearCombos {
   482  					x := third.a*(first.a+first.b) + third.b*(second.a+second.b)
   483  					cost := first.cost + second.cost + third.cost
   484  					old := c.MulRecipes[x]
   485  					if (old.Build == nil || cost < old.cost) && cost < mulCost {
   486  						c.MulRecipes[x] = mulRecipe{cost: cost, Build: func(m, v *Value) *Value {
   487  							v1 := first.build(m, v, v)
   488  							v2 := second.build(m, v, v)
   489  							return third.build(m, v1, v2)
   490  						}}
   491  					}
   492  				}
   493  			}
   494  		}
   495  
   496  		// E
   497  		for _, first := range linearCombos {
   498  			for _, second := range linearCombos {
   499  				for _, third := range linearCombos {
   500  					x := third.a*(second.a*(first.a+first.b)+second.b) + third.b
   501  					cost := first.cost + second.cost + third.cost
   502  					old := c.MulRecipes[x]
   503  					if (old.Build == nil || cost < old.cost) && cost < mulCost {
   504  						c.MulRecipes[x] = mulRecipe{cost: cost, Build: func(m, v *Value) *Value {
   505  							v1 := first.build(m, v, v)
   506  							v2 := second.build(m, v1, v)
   507  							return third.build(m, v2, v)
   508  						}}
   509  					}
   510  				}
   511  			}
   512  		}
   513  	}
   514  
   515  	// These cases should be handled specially by rewrite rules.
   516  	// (Otherwise v * 1 == (neg (neg v)))
   517  	delete(c.MulRecipes, 0)
   518  	delete(c.MulRecipes, 1)
   519  
   520  	// Currently:
   521  	// len(c.mulRecipes) == 5984 on arm64
   522  	//                       680 on amd64
   523  	//                      9738 on loong64
   524  	// This function takes ~2.5ms on arm64.
   525  	//println(len(c.mulRecipes))
   526  }
   527
View as plain text