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/types"
    12  	"cmd/internal/obj"
    13  	"cmd/internal/src"
    14  )
    15  
    16  // A Config holds readonly compilation information.
    17  // It is created once, early during compilation,
    18  // and shared across all compilations.
    19  type Config struct {
    20  	arch           string // "amd64", etc.
    21  	PtrSize        int64  // 4 or 8; copy of cmd/internal/sys.Arch.PtrSize
    22  	RegSize        int64  // 4 or 8; copy of cmd/internal/sys.Arch.RegSize
    23  	Types          Types
    24  	lowerBlock     blockRewriter  // block lowering function, first round
    25  	lowerValue     valueRewriter  // value lowering function, first round
    26  	lateLowerBlock blockRewriter  // block lowering function that needs to be run after the first round; only used on some architectures
    27  	lateLowerValue valueRewriter  // value lowering function that needs to be run after the first round; only used on some architectures
    28  	splitLoad      valueRewriter  // function for splitting merged load ops; only used on some architectures
    29  	registers      []Register     // machine registers
    30  	gpRegMask      regMask        // general purpose integer register mask
    31  	fpRegMask      regMask        // floating point register mask
    32  	fp32RegMask    regMask        // floating point register mask
    33  	fp64RegMask    regMask        // floating point register mask
    34  	specialRegMask regMask        // special register mask
    35  	intParamRegs   []int8         // register numbers of integer param (in/out) registers
    36  	floatParamRegs []int8         // register numbers of floating param (in/out) registers
    37  	ABI1           *abi.ABIConfig // "ABIInternal" under development // TODO change comment when this becomes current
    38  	ABI0           *abi.ABIConfig
    39  	FPReg          int8      // register number of frame pointer, -1 if not used
    40  	LinkReg        int8      // register number of link register if it is a general purpose register, -1 if not used
    41  	hasGReg        bool      // has hardware g register
    42  	ctxt           *obj.Link // Generic arch information
    43  	optimize       bool      // Do optimization
    44  	useAvg         bool      // Use optimizations that need Avg* operations
    45  	useHmul        bool      // Use optimizations that need Hmul* operations
    46  	SoftFloat      bool      //
    47  	Race           bool      // race detector enabled
    48  	BigEndian      bool      //
    49  	unalignedOK    bool      // Unaligned loads/stores are ok
    50  	haveBswap64    bool      // architecture implements Bswap64
    51  	haveBswap32    bool      // architecture implements Bswap32
    52  	haveBswap16    bool      // architecture implements Bswap16
    53  
    54  	// mulRecipes[x] = function to build v * x from v.
    55  	mulRecipes map[int64]mulRecipe
    56  }
    57  
    58  type mulRecipe struct {
    59  	cost  int
    60  	build func(*Value, *Value) *Value // build(m, v) returns v * x built at m.
    61  }
    62  
    63  type (
    64  	blockRewriter func(*Block) bool
    65  	valueRewriter func(*Value) bool
    66  )
    67  
    68  type Types struct {
    69  	Bool       *types.Type
    70  	Int8       *types.Type
    71  	Int16      *types.Type
    72  	Int32      *types.Type
    73  	Int64      *types.Type
    74  	UInt8      *types.Type
    75  	UInt16     *types.Type
    76  	UInt32     *types.Type
    77  	UInt64     *types.Type
    78  	Int        *types.Type
    79  	Float32    *types.Type
    80  	Float64    *types.Type
    81  	UInt       *types.Type
    82  	Uintptr    *types.Type
    83  	String     *types.Type
    84  	BytePtr    *types.Type // TODO: use unsafe.Pointer instead?
    85  	Int32Ptr   *types.Type
    86  	UInt32Ptr  *types.Type
    87  	IntPtr     *types.Type
    88  	UintptrPtr *types.Type
    89  	Float32Ptr *types.Type
    90  	Float64Ptr *types.Type
    91  	BytePtrPtr *types.Type
    92  }
    93  
    94  // NewTypes creates and populates a Types.
    95  func NewTypes() *Types {
    96  	t := new(Types)
    97  	t.SetTypPtrs()
    98  	return t
    99  }
   100  
   101  // SetTypPtrs populates t.
   102  func (t *Types) SetTypPtrs() {
   103  	t.Bool = types.Types[types.TBOOL]
   104  	t.Int8 = types.Types[types.TINT8]
   105  	t.Int16 = types.Types[types.TINT16]
   106  	t.Int32 = types.Types[types.TINT32]
   107  	t.Int64 = types.Types[types.TINT64]
   108  	t.UInt8 = types.Types[types.TUINT8]
   109  	t.UInt16 = types.Types[types.TUINT16]
   110  	t.UInt32 = types.Types[types.TUINT32]
   111  	t.UInt64 = types.Types[types.TUINT64]
   112  	t.Int = types.Types[types.TINT]
   113  	t.Float32 = types.Types[types.TFLOAT32]
   114  	t.Float64 = types.Types[types.TFLOAT64]
   115  	t.UInt = types.Types[types.TUINT]
   116  	t.Uintptr = types.Types[types.TUINTPTR]
   117  	t.String = types.Types[types.TSTRING]
   118  	t.BytePtr = types.NewPtr(types.Types[types.TUINT8])
   119  	t.Int32Ptr = types.NewPtr(types.Types[types.TINT32])
   120  	t.UInt32Ptr = types.NewPtr(types.Types[types.TUINT32])
   121  	t.IntPtr = types.NewPtr(types.Types[types.TINT])
   122  	t.UintptrPtr = types.NewPtr(types.Types[types.TUINTPTR])
   123  	t.Float32Ptr = types.NewPtr(types.Types[types.TFLOAT32])
   124  	t.Float64Ptr = types.NewPtr(types.Types[types.TFLOAT64])
   125  	t.BytePtrPtr = types.NewPtr(types.NewPtr(types.Types[types.TUINT8]))
   126  }
   127  
   128  type Logger interface {
   129  	// Logf logs a message from the compiler.
   130  	Logf(string, ...interface{})
   131  
   132  	// Log reports whether logging is not a no-op
   133  	// some logging calls account for more than a few heap allocations.
   134  	Log() bool
   135  
   136  	// Fatalf reports a compiler error and exits.
   137  	Fatalf(pos src.XPos, msg string, args ...interface{})
   138  
   139  	// Warnl writes compiler messages in the form expected by "errorcheck" tests
   140  	Warnl(pos src.XPos, fmt_ string, args ...interface{})
   141  
   142  	// Forwards the Debug flags from gc
   143  	Debug_checknil() bool
   144  }
   145  
   146  type Frontend interface {
   147  	Logger
   148  
   149  	// StringData returns a symbol pointing to the given string's contents.
   150  	StringData(string) *obj.LSym
   151  
   152  	// Given the name for a compound type, returns the name we should use
   153  	// for the parts of that compound type.
   154  	SplitSlot(parent *LocalSlot, suffix string, offset int64, t *types.Type) LocalSlot
   155  
   156  	// Syslook returns a symbol of the runtime function/variable with the
   157  	// given name.
   158  	Syslook(string) *obj.LSym
   159  
   160  	// UseWriteBarrier reports whether write barrier is enabled
   161  	UseWriteBarrier() bool
   162  
   163  	// Func returns the ir.Func of the function being compiled.
   164  	Func() *ir.Func
   165  }
   166  
   167  // NewConfig returns a new configuration object for the given architecture.
   168  func NewConfig(arch string, types Types, ctxt *obj.Link, optimize, softfloat bool) *Config {
   169  	c := &Config{arch: arch, Types: types}
   170  	c.useAvg = true
   171  	c.useHmul = true
   172  	switch arch {
   173  	case "amd64":
   174  		c.PtrSize = 8
   175  		c.RegSize = 8
   176  		c.lowerBlock = rewriteBlockAMD64
   177  		c.lowerValue = rewriteValueAMD64
   178  		c.lateLowerBlock = rewriteBlockAMD64latelower
   179  		c.lateLowerValue = rewriteValueAMD64latelower
   180  		c.splitLoad = rewriteValueAMD64splitload
   181  		c.registers = registersAMD64[:]
   182  		c.gpRegMask = gpRegMaskAMD64
   183  		c.fpRegMask = fpRegMaskAMD64
   184  		c.specialRegMask = specialRegMaskAMD64
   185  		c.intParamRegs = paramIntRegAMD64
   186  		c.floatParamRegs = paramFloatRegAMD64
   187  		c.FPReg = framepointerRegAMD64
   188  		c.LinkReg = linkRegAMD64
   189  		c.hasGReg = true
   190  		c.unalignedOK = true
   191  		c.haveBswap64 = true
   192  		c.haveBswap32 = true
   193  		c.haveBswap16 = true
   194  	case "386":
   195  		c.PtrSize = 4
   196  		c.RegSize = 4
   197  		c.lowerBlock = rewriteBlock386
   198  		c.lowerValue = rewriteValue386
   199  		c.splitLoad = rewriteValue386splitload
   200  		c.registers = registers386[:]
   201  		c.gpRegMask = gpRegMask386
   202  		c.fpRegMask = fpRegMask386
   203  		c.FPReg = framepointerReg386
   204  		c.LinkReg = linkReg386
   205  		c.hasGReg = false
   206  		c.unalignedOK = true
   207  		c.haveBswap32 = true
   208  		c.haveBswap16 = true
   209  	case "arm":
   210  		c.PtrSize = 4
   211  		c.RegSize = 4
   212  		c.lowerBlock = rewriteBlockARM
   213  		c.lowerValue = rewriteValueARM
   214  		c.registers = registersARM[:]
   215  		c.gpRegMask = gpRegMaskARM
   216  		c.fpRegMask = fpRegMaskARM
   217  		c.FPReg = framepointerRegARM
   218  		c.LinkReg = linkRegARM
   219  		c.hasGReg = true
   220  	case "arm64":
   221  		c.PtrSize = 8
   222  		c.RegSize = 8
   223  		c.lowerBlock = rewriteBlockARM64
   224  		c.lowerValue = rewriteValueARM64
   225  		c.lateLowerBlock = rewriteBlockARM64latelower
   226  		c.lateLowerValue = rewriteValueARM64latelower
   227  		c.registers = registersARM64[:]
   228  		c.gpRegMask = gpRegMaskARM64
   229  		c.fpRegMask = fpRegMaskARM64
   230  		c.intParamRegs = paramIntRegARM64
   231  		c.floatParamRegs = paramFloatRegARM64
   232  		c.FPReg = framepointerRegARM64
   233  		c.LinkReg = linkRegARM64
   234  		c.hasGReg = true
   235  		c.unalignedOK = true
   236  		c.haveBswap64 = true
   237  		c.haveBswap32 = true
   238  		c.haveBswap16 = true
   239  	case "ppc64":
   240  		c.BigEndian = true
   241  		fallthrough
   242  	case "ppc64le":
   243  		c.PtrSize = 8
   244  		c.RegSize = 8
   245  		c.lowerBlock = rewriteBlockPPC64
   246  		c.lowerValue = rewriteValuePPC64
   247  		c.lateLowerBlock = rewriteBlockPPC64latelower
   248  		c.lateLowerValue = rewriteValuePPC64latelower
   249  		c.registers = registersPPC64[:]
   250  		c.gpRegMask = gpRegMaskPPC64
   251  		c.fpRegMask = fpRegMaskPPC64
   252  		c.specialRegMask = specialRegMaskPPC64
   253  		c.intParamRegs = paramIntRegPPC64
   254  		c.floatParamRegs = paramFloatRegPPC64
   255  		c.FPReg = framepointerRegPPC64
   256  		c.LinkReg = linkRegPPC64
   257  		c.hasGReg = true
   258  		c.unalignedOK = true
   259  		// Note: ppc64 has register bswap ops only when GOPPC64>=10.
   260  		// But it has bswap+load and bswap+store ops for all ppc64 variants.
   261  		// That is the sense we're using them here - they are only used
   262  		// in contexts where they can be merged with a load or store.
   263  		c.haveBswap64 = true
   264  		c.haveBswap32 = true
   265  		c.haveBswap16 = true
   266  	case "mips64":
   267  		c.BigEndian = true
   268  		fallthrough
   269  	case "mips64le":
   270  		c.PtrSize = 8
   271  		c.RegSize = 8
   272  		c.lowerBlock = rewriteBlockMIPS64
   273  		c.lowerValue = rewriteValueMIPS64
   274  		c.registers = registersMIPS64[:]
   275  		c.gpRegMask = gpRegMaskMIPS64
   276  		c.fpRegMask = fpRegMaskMIPS64
   277  		c.specialRegMask = specialRegMaskMIPS64
   278  		c.FPReg = framepointerRegMIPS64
   279  		c.LinkReg = linkRegMIPS64
   280  		c.hasGReg = true
   281  	case "loong64":
   282  		c.PtrSize = 8
   283  		c.RegSize = 8
   284  		c.lowerBlock = rewriteBlockLOONG64
   285  		c.lowerValue = rewriteValueLOONG64
   286  		c.lateLowerBlock = rewriteBlockLOONG64latelower
   287  		c.lateLowerValue = rewriteValueLOONG64latelower
   288  		c.registers = registersLOONG64[:]
   289  		c.gpRegMask = gpRegMaskLOONG64
   290  		c.fpRegMask = fpRegMaskLOONG64
   291  		c.intParamRegs = paramIntRegLOONG64
   292  		c.floatParamRegs = paramFloatRegLOONG64
   293  		c.FPReg = framepointerRegLOONG64
   294  		c.LinkReg = linkRegLOONG64
   295  		c.hasGReg = true
   296  		c.unalignedOK = true
   297  	case "s390x":
   298  		c.PtrSize = 8
   299  		c.RegSize = 8
   300  		c.lowerBlock = rewriteBlockS390X
   301  		c.lowerValue = rewriteValueS390X
   302  		c.registers = registersS390X[:]
   303  		c.gpRegMask = gpRegMaskS390X
   304  		c.fpRegMask = fpRegMaskS390X
   305  		c.FPReg = framepointerRegS390X
   306  		c.LinkReg = linkRegS390X
   307  		c.hasGReg = true
   308  		c.BigEndian = true
   309  		c.unalignedOK = true
   310  		c.haveBswap64 = true
   311  		c.haveBswap32 = true
   312  		c.haveBswap16 = true // only for loads&stores, see ppc64 comment
   313  	case "mips":
   314  		c.BigEndian = true
   315  		fallthrough
   316  	case "mipsle":
   317  		c.PtrSize = 4
   318  		c.RegSize = 4
   319  		c.lowerBlock = rewriteBlockMIPS
   320  		c.lowerValue = rewriteValueMIPS
   321  		c.registers = registersMIPS[:]
   322  		c.gpRegMask = gpRegMaskMIPS
   323  		c.fpRegMask = fpRegMaskMIPS
   324  		c.specialRegMask = specialRegMaskMIPS
   325  		c.FPReg = framepointerRegMIPS
   326  		c.LinkReg = linkRegMIPS
   327  		c.hasGReg = true
   328  	case "riscv64":
   329  		c.PtrSize = 8
   330  		c.RegSize = 8
   331  		c.lowerBlock = rewriteBlockRISCV64
   332  		c.lowerValue = rewriteValueRISCV64
   333  		c.lateLowerBlock = rewriteBlockRISCV64latelower
   334  		c.lateLowerValue = rewriteValueRISCV64latelower
   335  		c.registers = registersRISCV64[:]
   336  		c.gpRegMask = gpRegMaskRISCV64
   337  		c.fpRegMask = fpRegMaskRISCV64
   338  		c.intParamRegs = paramIntRegRISCV64
   339  		c.floatParamRegs = paramFloatRegRISCV64
   340  		c.FPReg = framepointerRegRISCV64
   341  		c.hasGReg = true
   342  	case "wasm":
   343  		c.PtrSize = 8
   344  		c.RegSize = 8
   345  		c.lowerBlock = rewriteBlockWasm
   346  		c.lowerValue = rewriteValueWasm
   347  		c.registers = registersWasm[:]
   348  		c.gpRegMask = gpRegMaskWasm
   349  		c.fpRegMask = fpRegMaskWasm
   350  		c.fp32RegMask = fp32RegMaskWasm
   351  		c.fp64RegMask = fp64RegMaskWasm
   352  		c.FPReg = framepointerRegWasm
   353  		c.LinkReg = linkRegWasm
   354  		c.hasGReg = true
   355  		c.useAvg = false
   356  		c.useHmul = false
   357  	default:
   358  		ctxt.Diag("arch %s not implemented", arch)
   359  	}
   360  	c.ctxt = ctxt
   361  	c.optimize = optimize
   362  	c.SoftFloat = softfloat
   363  	if softfloat {
   364  		c.floatParamRegs = nil // no FP registers in softfloat mode
   365  	}
   366  
   367  	c.ABI0 = abi.NewABIConfig(0, 0, ctxt.Arch.FixedFrameSize, 0)
   368  	c.ABI1 = abi.NewABIConfig(len(c.intParamRegs), len(c.floatParamRegs), ctxt.Arch.FixedFrameSize, 1)
   369  
   370  	if ctxt.Flag_shared {
   371  		// LoweredWB is secretly a CALL and CALLs on 386 in
   372  		// shared mode get rewritten by obj6.go to go through
   373  		// the GOT, which clobbers BX.
   374  		opcodeTable[Op386LoweredWB].reg.clobbers |= 1 << 3 // BX
   375  	}
   376  
   377  	c.buildRecipes(arch)
   378  
   379  	return c
   380  }
   381  
   382  func (c *Config) Ctxt() *obj.Link { return c.ctxt }
   383  
   384  func (c *Config) haveByteSwap(size int64) bool {
   385  	switch size {
   386  	case 8:
   387  		return c.haveBswap64
   388  	case 4:
   389  		return c.haveBswap32
   390  	case 2:
   391  		return c.haveBswap16
   392  	default:
   393  		base.Fatalf("bad size %d\n", size)
   394  		return false
   395  	}
   396  }
   397  
   398  func (c *Config) buildRecipes(arch string) {
   399  	// Information for strength-reducing multiplies.
   400  	type linearCombo struct {
   401  		// we can compute a*x+b*y in one instruction
   402  		a, b int64
   403  		// cost, in arbitrary units (tenths of cycles, usually)
   404  		cost int
   405  		// builds SSA value for a*x+b*y. Use the position
   406  		// information from m.
   407  		build func(m, x, y *Value) *Value
   408  	}
   409  
   410  	// List all the linear combination instructions we have.
   411  	var linearCombos []linearCombo
   412  	r := func(a, b int64, cost int, build func(m, x, y *Value) *Value) {
   413  		linearCombos = append(linearCombos, linearCombo{a: a, b: b, cost: cost, build: build})
   414  	}
   415  	var mulCost int
   416  	switch arch {
   417  	case "amd64":
   418  		// Assumes that the following costs from https://gmplib.org/~tege/x86-timing.pdf:
   419  		//    1 - addq, shlq, leaq, negq, subq
   420  		//    3 - imulq
   421  		// These costs limit the rewrites to two instructions.
   422  		// Operations which have to happen in place (and thus
   423  		// may require a reg-reg move) score slightly higher.
   424  		mulCost = 30
   425  		// add
   426  		r(1, 1, 10,
   427  			func(m, x, y *Value) *Value {
   428  				v := m.Block.NewValue2(m.Pos, OpAMD64ADDQ, m.Type, x, y)
   429  				if m.Type.Size() == 4 {
   430  					v.Op = OpAMD64ADDL
   431  				}
   432  				return v
   433  			})
   434  		// neg
   435  		r(-1, 0, 11,
   436  			func(m, x, y *Value) *Value {
   437  				v := m.Block.NewValue1(m.Pos, OpAMD64NEGQ, m.Type, x)
   438  				if m.Type.Size() == 4 {
   439  					v.Op = OpAMD64NEGL
   440  				}
   441  				return v
   442  			})
   443  		// sub
   444  		r(1, -1, 11,
   445  			func(m, x, y *Value) *Value {
   446  				v := m.Block.NewValue2(m.Pos, OpAMD64SUBQ, m.Type, x, y)
   447  				if m.Type.Size() == 4 {
   448  					v.Op = OpAMD64SUBL
   449  				}
   450  				return v
   451  			})
   452  		// lea
   453  		r(1, 2, 10,
   454  			func(m, x, y *Value) *Value {
   455  				v := m.Block.NewValue2(m.Pos, OpAMD64LEAQ2, m.Type, x, y)
   456  				if m.Type.Size() == 4 {
   457  					v.Op = OpAMD64LEAL2
   458  				}
   459  				return v
   460  			})
   461  		r(1, 4, 10,
   462  			func(m, x, y *Value) *Value {
   463  				v := m.Block.NewValue2(m.Pos, OpAMD64LEAQ4, m.Type, x, y)
   464  				if m.Type.Size() == 4 {
   465  					v.Op = OpAMD64LEAL4
   466  				}
   467  				return v
   468  			})
   469  		r(1, 8, 10,
   470  			func(m, x, y *Value) *Value {
   471  				v := m.Block.NewValue2(m.Pos, OpAMD64LEAQ8, m.Type, x, y)
   472  				if m.Type.Size() == 4 {
   473  					v.Op = OpAMD64LEAL8
   474  				}
   475  				return v
   476  			})
   477  		// regular shifts
   478  		for i := 2; i < 64; i++ {
   479  			r(1<<i, 0, 11,
   480  				func(m, x, y *Value) *Value {
   481  					v := m.Block.NewValue1I(m.Pos, OpAMD64SHLQconst, m.Type, int64(i), x)
   482  					if m.Type.Size() == 4 {
   483  						v.Op = OpAMD64SHLLconst
   484  					}
   485  					return v
   486  				})
   487  		}
   488  
   489  	case "arm64":
   490  		// Rationale (for M2 ultra):
   491  		// - multiply is 3 cycles.
   492  		// - add/neg/sub/shift are 1 cycle.
   493  		// - add/neg/sub+shiftLL are 2 cycles.
   494  		// We break ties against the multiply because using a
   495  		// multiply also needs to load the constant into a register.
   496  		// (It's 3 cycles and 2 instructions either way, but the
   497  		// linear combo one might use 1 less register.)
   498  		// The multiply constant might get lifted out of a loop though. Hmm....
   499  		// Other arm64 chips have different tradeoffs.
   500  		// Some chip's add+shift instructions are 1 cycle for shifts up to 4
   501  		// and 2 cycles for shifts bigger than 4. So weight the larger shifts
   502  		// a bit more.
   503  		// TODO: figure out a happy medium.
   504  		mulCost = 35
   505  		// add
   506  		r(1, 1, 10,
   507  			func(m, x, y *Value) *Value {
   508  				return m.Block.NewValue2(m.Pos, OpARM64ADD, m.Type, x, y)
   509  			})
   510  		// neg
   511  		r(-1, 0, 10,
   512  			func(m, x, y *Value) *Value {
   513  				return m.Block.NewValue1(m.Pos, OpARM64NEG, m.Type, x)
   514  			})
   515  		// sub
   516  		r(1, -1, 10,
   517  			func(m, x, y *Value) *Value {
   518  				return m.Block.NewValue2(m.Pos, OpARM64SUB, m.Type, x, y)
   519  			})
   520  		// regular shifts
   521  		for i := 1; i < 64; i++ {
   522  			c := 10
   523  			if i == 1 {
   524  				// Prefer x<<1 over x+x.
   525  				// Note that we eventually reverse this decision in ARM64latelower.rules,
   526  				// but this makes shift combining rules in ARM64.rules simpler.
   527  				c--
   528  			}
   529  			r(1<<i, 0, c,
   530  				func(m, x, y *Value) *Value {
   531  					return m.Block.NewValue1I(m.Pos, OpARM64SLLconst, m.Type, int64(i), x)
   532  				})
   533  		}
   534  		// ADDshiftLL
   535  		for i := 1; i < 64; i++ {
   536  			c := 20
   537  			if i > 4 {
   538  				c++
   539  			}
   540  			r(1, 1<<i, c,
   541  				func(m, x, y *Value) *Value {
   542  					return m.Block.NewValue2I(m.Pos, OpARM64ADDshiftLL, m.Type, int64(i), x, y)
   543  				})
   544  		}
   545  		// NEGshiftLL
   546  		for i := 1; i < 64; i++ {
   547  			c := 20
   548  			if i > 4 {
   549  				c++
   550  			}
   551  			r(-1<<i, 0, c,
   552  				func(m, x, y *Value) *Value {
   553  					return m.Block.NewValue1I(m.Pos, OpARM64NEGshiftLL, m.Type, int64(i), x)
   554  				})
   555  		}
   556  		// SUBshiftLL
   557  		for i := 1; i < 64; i++ {
   558  			c := 20
   559  			if i > 4 {
   560  				c++
   561  			}
   562  			r(1, -1<<i, c,
   563  				func(m, x, y *Value) *Value {
   564  					return m.Block.NewValue2I(m.Pos, OpARM64SUBshiftLL, m.Type, int64(i), x, y)
   565  				})
   566  		}
   567  	case "loong64":
   568  		// - multiply is 4 cycles.
   569  		// - add/sub/shift are 1 cycle.
   570  		// On loong64, using a multiply also needs to load the constant into a register.
   571  		// TODO: figure out a happy medium.
   572  		mulCost = 45
   573  
   574  		// add
   575  		r(1, 1, 10,
   576  			func(m, x, y *Value) *Value {
   577  				return m.Block.NewValue2(m.Pos, OpLOONG64ADDV, m.Type, x, y)
   578  			})
   579  		// neg
   580  		r(-1, 0, 10,
   581  			func(m, x, y *Value) *Value {
   582  				return m.Block.NewValue1(m.Pos, OpLOONG64NEGV, m.Type, x)
   583  			})
   584  		// sub
   585  		r(1, -1, 10,
   586  			func(m, x, y *Value) *Value {
   587  				return m.Block.NewValue2(m.Pos, OpLOONG64SUBV, m.Type, x, y)
   588  			})
   589  
   590  		// regular shifts
   591  		for i := 1; i < 64; i++ {
   592  			c := 10
   593  			if i == 1 {
   594  				// Prefer x<<1 over x+x.
   595  				// Note that we eventually reverse this decision in LOONG64latelower.rules,
   596  				// but this makes shift combining rules in LOONG64.rules simpler.
   597  				c--
   598  			}
   599  			r(1<<i, 0, c,
   600  				func(m, x, y *Value) *Value {
   601  					return m.Block.NewValue1I(m.Pos, OpLOONG64SLLVconst, m.Type, int64(i), x)
   602  				})
   603  		}
   604  	}
   605  
   606  	c.mulRecipes = map[int64]mulRecipe{}
   607  
   608  	// Single-instruction recipes.
   609  	// The only option for the input value(s) is v.
   610  	for _, combo := range linearCombos {
   611  		x := combo.a + combo.b
   612  		cost := combo.cost
   613  		old := c.mulRecipes[x]
   614  		if (old.build == nil || cost < old.cost) && cost < mulCost {
   615  			c.mulRecipes[x] = mulRecipe{cost: cost, build: func(m, v *Value) *Value {
   616  				return combo.build(m, v, v)
   617  			}}
   618  		}
   619  	}
   620  	// Two-instruction recipes.
   621  	// A: Both of the outer's inputs are from the same single-instruction recipe.
   622  	// B: First input is v and the second is from a single-instruction recipe.
   623  	// C: Second input is v and the first is from a single-instruction recipe.
   624  	// A is slightly preferred because it often needs 1 less register, so it
   625  	// goes first.
   626  
   627  	// A
   628  	for _, inner := range linearCombos {
   629  		for _, outer := range linearCombos {
   630  			x := (inner.a + inner.b) * (outer.a + outer.b)
   631  			cost := inner.cost + outer.cost
   632  			old := c.mulRecipes[x]
   633  			if (old.build == nil || cost < old.cost) && cost < mulCost {
   634  				c.mulRecipes[x] = mulRecipe{cost: cost, build: func(m, v *Value) *Value {
   635  					v = inner.build(m, v, v)
   636  					return outer.build(m, v, v)
   637  				}}
   638  			}
   639  		}
   640  	}
   641  
   642  	// B
   643  	for _, inner := range linearCombos {
   644  		for _, outer := range linearCombos {
   645  			x := outer.a + outer.b*(inner.a+inner.b)
   646  			cost := inner.cost + outer.cost
   647  			old := c.mulRecipes[x]
   648  			if (old.build == nil || cost < old.cost) && cost < mulCost {
   649  				c.mulRecipes[x] = mulRecipe{cost: cost, build: func(m, v *Value) *Value {
   650  					return outer.build(m, v, inner.build(m, v, v))
   651  				}}
   652  			}
   653  		}
   654  	}
   655  
   656  	// C
   657  	for _, inner := range linearCombos {
   658  		for _, outer := range linearCombos {
   659  			x := outer.a*(inner.a+inner.b) + outer.b
   660  			cost := inner.cost + outer.cost
   661  			old := c.mulRecipes[x]
   662  			if (old.build == nil || cost < old.cost) && cost < mulCost {
   663  				c.mulRecipes[x] = mulRecipe{cost: cost, build: func(m, v *Value) *Value {
   664  					return outer.build(m, inner.build(m, v, v), v)
   665  				}}
   666  			}
   667  		}
   668  	}
   669  
   670  	// Currently we only process 3 linear combination instructions for loong64.
   671  	if arch == "loong64" {
   672  		// Three-instruction recipes.
   673  		// D: The first and the second are all single-instruction recipes, and they are also the third's inputs.
   674  		// E: The first single-instruction is the second's input, and the second is the third's input.
   675  
   676  		// D
   677  		for _, first := range linearCombos {
   678  			for _, second := range linearCombos {
   679  				for _, third := range linearCombos {
   680  					x := third.a*(first.a+first.b) + third.b*(second.a+second.b)
   681  					cost := first.cost + second.cost + third.cost
   682  					old := c.mulRecipes[x]
   683  					if (old.build == nil || cost < old.cost) && cost < mulCost {
   684  						c.mulRecipes[x] = mulRecipe{cost: cost, build: func(m, v *Value) *Value {
   685  							v1 := first.build(m, v, v)
   686  							v2 := second.build(m, v, v)
   687  							return third.build(m, v1, v2)
   688  						}}
   689  					}
   690  				}
   691  			}
   692  		}
   693  
   694  		// E
   695  		for _, first := range linearCombos {
   696  			for _, second := range linearCombos {
   697  				for _, third := range linearCombos {
   698  					x := third.a*(second.a*(first.a+first.b)+second.b) + third.b
   699  					cost := first.cost + second.cost + third.cost
   700  					old := c.mulRecipes[x]
   701  					if (old.build == nil || cost < old.cost) && cost < mulCost {
   702  						c.mulRecipes[x] = mulRecipe{cost: cost, build: func(m, v *Value) *Value {
   703  							v1 := first.build(m, v, v)
   704  							v2 := second.build(m, v1, v)
   705  							return third.build(m, v2, v)
   706  						}}
   707  					}
   708  				}
   709  			}
   710  		}
   711  	}
   712  
   713  	// These cases should be handled specially by rewrite rules.
   714  	// (Otherwise v * 1 == (neg (neg v)))
   715  	delete(c.mulRecipes, 0)
   716  	delete(c.mulRecipes, 1)
   717  
   718  	// Currently:
   719  	// len(c.mulRecipes) == 5984 on arm64
   720  	//                       680 on amd64
   721  	//                      5984 on loong64
   722  	// This function takes ~2.5ms on arm64.
   723  	//println(len(c.mulRecipes))
   724  }
   725
View as plain text