Source file src/cmd/compile/internal/ssa/expand_calls.go

     1  // Copyright 2020 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/src"
    13  	"fmt"
    14  )
    15  
    16  func postExpandCallsDecompose(f *Func) {
    17  	decomposeUser(f)    // redo user decompose to cleanup after expand calls
    18  	decomposeBuiltIn(f) // handles both regular decomposition and cleanup.
    19  }
    20  
    21  func expandCalls(f *Func) {
    22  	// Convert each aggregate arg to a call into "dismantle aggregate, store/pass parts"
    23  	// Convert each aggregate result from a call into "assemble aggregate from parts"
    24  	// Convert each multivalue exit into "dismantle aggregate, store/return parts"
    25  	// Convert incoming aggregate arg into assembly of parts.
    26  	// Feed modified AST to decompose.
    27  
    28  	sp, _ := f.spSb()
    29  
    30  	x := &expandState{
    31  		f:               f,
    32  		debug:           f.pass.debug,
    33  		regSize:         f.Config.RegSize,
    34  		sp:              sp,
    35  		typs:            &f.Config.Types,
    36  		wideSelects:     make(map[*Value]*Value),
    37  		commonArgs:      make(map[selKey]*Value),
    38  		commonSelectors: make(map[selKey]*Value),
    39  		memForCall:      make(map[ID]*Value),
    40  	}
    41  
    42  	// For 32-bit, need to deal with decomposition of 64-bit integers, which depends on endianness.
    43  	if f.Config.BigEndian {
    44  		x.firstOp = OpInt64Hi
    45  		x.secondOp = OpInt64Lo
    46  		x.firstType = x.typs.Int32
    47  		x.secondType = x.typs.UInt32
    48  	} else {
    49  		x.firstOp = OpInt64Lo
    50  		x.secondOp = OpInt64Hi
    51  		x.firstType = x.typs.UInt32
    52  		x.secondType = x.typs.Int32
    53  	}
    54  
    55  	// Defer select processing until after all calls and selects are seen.
    56  	var selects []*Value
    57  	var calls []*Value
    58  	var args []*Value
    59  	var exitBlocks []*Block
    60  
    61  	var m0 *Value
    62  
    63  	// Accumulate lists of calls, args, selects, and exit blocks to process,
    64  	// note "wide" selects consumed by stores,
    65  	// rewrite mem for each call,
    66  	// rewrite each OpSelectNAddr.
    67  	for _, b := range f.Blocks {
    68  		for _, v := range b.Values {
    69  			switch v.Op {
    70  			case OpInitMem:
    71  				m0 = v
    72  
    73  			case OpClosureLECall, OpInterLECall, OpStaticLECall, OpTailLECall:
    74  				calls = append(calls, v)
    75  
    76  			case OpArg:
    77  				args = append(args, v)
    78  
    79  			case OpStore:
    80  				if a := v.Args[1]; a.Op == OpSelectN && !CanSSA(a.Type) {
    81  					if a.Uses > 1 {
    82  						panic(fmt.Errorf("Saw double use of wide SelectN %s operand of Store %s",
    83  							a.LongString(), v.LongString()))
    84  					}
    85  					x.wideSelects[a] = v
    86  				}
    87  
    88  			case OpSelectN:
    89  				if v.Type == types.TypeMem {
    90  					// rewrite the mem selector in place
    91  					call := v.Args[0]
    92  					aux := call.Aux.(*AuxCall)
    93  					mem := x.memForCall[call.ID]
    94  					if mem == nil {
    95  						v.AuxInt = int64(aux.abiInfo.OutRegistersUsed())
    96  						x.memForCall[call.ID] = v
    97  					} else {
    98  						panic(fmt.Errorf("Saw two memories for call %v, %v and %v", call, mem, v))
    99  					}
   100  				} else {
   101  					selects = append(selects, v)
   102  				}
   103  
   104  			case OpSelectNAddr:
   105  				call := v.Args[0]
   106  				which := v.AuxInt
   107  				aux := call.Aux.(*AuxCall)
   108  				pt := v.Type
   109  				off := x.offsetFrom(x.f.Entry, x.sp, aux.OffsetOfResult(which), pt)
   110  				v.copyOf(off)
   111  			}
   112  		}
   113  
   114  		// rewrite function results from an exit block
   115  		// values returned by function need to be split out into registers.
   116  		if isBlockMultiValueExit(b) {
   117  			exitBlocks = append(exitBlocks, b)
   118  		}
   119  	}
   120  
   121  	// Convert each aggregate arg into Make of its parts (and so on, to primitive types)
   122  	for _, v := range args {
   123  		var rc registerCursor
   124  		a := x.prAssignForArg(v)
   125  		aux := x.f.OwnAux
   126  		regs := a.Registers
   127  		var offset int64
   128  		if len(regs) == 0 {
   129  			offset = a.FrameOffset(aux.abiInfo)
   130  		}
   131  		auxBase := x.offsetFrom(x.f.Entry, x.sp, offset, types.NewPtr(v.Type))
   132  		rc.init(regs, aux.abiInfo, nil, auxBase, 0)
   133  		x.rewriteSelectOrArg(f.Entry.Pos, f.Entry, v, v, m0, v.Type, rc)
   134  	}
   135  
   136  	// Rewrite selects of results (which may be aggregates) into make-aggregates of register/memory-targeted selects
   137  	for _, v := range selects {
   138  		if v.Op == OpInvalid {
   139  			continue
   140  		}
   141  
   142  		call := v.Args[0]
   143  		aux := call.Aux.(*AuxCall)
   144  		mem := x.memForCall[call.ID]
   145  		if mem == nil {
   146  			mem = call.Block.NewValue1I(call.Pos, OpSelectN, types.TypeMem, int64(aux.abiInfo.OutRegistersUsed()), call)
   147  			x.memForCall[call.ID] = mem
   148  		}
   149  
   150  		i := v.AuxInt
   151  		regs := aux.RegsOfResult(i)
   152  
   153  		// If this select cannot fit into SSA and is stored, either disaggregate to register stores, or mem-mem move.
   154  		if store := x.wideSelects[v]; store != nil {
   155  			// Use the mem that comes from the store operation.
   156  			storeAddr := store.Args[0]
   157  			mem := store.Args[2]
   158  			if len(regs) > 0 {
   159  				// Cannot do a rewrite that builds up a result from pieces; instead, copy pieces to the store operation.
   160  				var rc registerCursor
   161  				rc.init(regs, aux.abiInfo, nil, storeAddr, 0)
   162  				mem = x.rewriteWideSelectToStores(call.Pos, call.Block, v, mem, v.Type, rc)
   163  				store.copyOf(mem)
   164  			} else {
   165  				// Move directly from AuxBase to store target; rewrite the store instruction.
   166  				offset := aux.OffsetOfResult(i)
   167  				auxBase := x.offsetFrom(x.f.Entry, x.sp, offset, types.NewPtr(v.Type))
   168  				// was Store dst, v, mem
   169  				// now Move dst, auxBase, mem
   170  				move := store.Block.NewValue3A(store.Pos, OpMove, types.TypeMem, v.Type, storeAddr, auxBase, mem)
   171  				move.AuxInt = v.Type.Size()
   172  				store.copyOf(move)
   173  			}
   174  			continue
   175  		}
   176  
   177  		var auxBase *Value
   178  		if len(regs) == 0 {
   179  			offset := aux.OffsetOfResult(i)
   180  			auxBase = x.offsetFrom(x.f.Entry, x.sp, offset, types.NewPtr(v.Type))
   181  		}
   182  		var rc registerCursor
   183  		rc.init(regs, aux.abiInfo, nil, auxBase, 0)
   184  		x.rewriteSelectOrArg(call.Pos, call.Block, v, v, mem, v.Type, rc)
   185  	}
   186  
   187  	rewriteCall := func(v *Value, newOp Op, argStart int) {
   188  		// Break aggregate args passed to call into smaller pieces.
   189  		x.rewriteCallArgs(v, argStart)
   190  		v.Op = newOp
   191  		rts := abi.RegisterTypes(v.Aux.(*AuxCall).abiInfo.OutParams())
   192  		v.Type = types.NewResults(append(rts, types.TypeMem))
   193  	}
   194  
   195  	// Rewrite calls
   196  	for _, v := range calls {
   197  		switch v.Op {
   198  		case OpStaticLECall:
   199  			rewriteCall(v, OpStaticCall, 0)
   200  		case OpTailLECall:
   201  			rewriteCall(v, OpTailCall, 0)
   202  		case OpClosureLECall:
   203  			rewriteCall(v, OpClosureCall, 2)
   204  		case OpInterLECall:
   205  			rewriteCall(v, OpInterCall, 1)
   206  		}
   207  	}
   208  
   209  	// Rewrite results from exit blocks
   210  	for _, b := range exitBlocks {
   211  		v := b.Controls[0]
   212  		x.rewriteFuncResults(v, b, f.OwnAux)
   213  		b.SetControl(v)
   214  	}
   215  
   216  }
   217  
   218  func (x *expandState) rewriteFuncResults(v *Value, b *Block, aux *AuxCall) {
   219  	// This is very similar to rewriteCallArgs
   220  	// differences:
   221  	// firstArg + preArgs
   222  	// sp vs auxBase
   223  
   224  	m0 := v.MemoryArg()
   225  	mem := m0
   226  
   227  	allResults := []*Value{}
   228  	var oldArgs []*Value
   229  	argsWithoutMem := v.Args[:len(v.Args)-1]
   230  
   231  	for j, a := range argsWithoutMem {
   232  		oldArgs = append(oldArgs, a)
   233  		i := int64(j)
   234  		auxType := aux.TypeOfResult(i)
   235  		auxBase := b.NewValue2A(v.Pos, OpLocalAddr, types.NewPtr(auxType), aux.NameOfResult(i), x.sp, mem)
   236  		auxOffset := int64(0)
   237  		aRegs := aux.RegsOfResult(int64(j))
   238  		if a.Op == OpDereference {
   239  			a.Op = OpLoad
   240  		}
   241  		var rc registerCursor
   242  		var result *[]*Value
   243  		if len(aRegs) > 0 {
   244  			result = &allResults
   245  		} else {
   246  			if a.Op == OpLoad && a.Args[0].Op == OpLocalAddr {
   247  				addr := a.Args[0]
   248  				if addr.MemoryArg() == a.MemoryArg() && addr.Aux == aux.NameOfResult(i) {
   249  					continue // Self move to output parameter
   250  				}
   251  			}
   252  		}
   253  		rc.init(aRegs, aux.abiInfo, result, auxBase, auxOffset)
   254  		mem = x.decomposeAsNecessary(v.Pos, b, a, mem, rc)
   255  	}
   256  	v.resetArgs()
   257  	v.AddArgs(allResults...)
   258  	v.AddArg(mem)
   259  	for _, a := range oldArgs {
   260  		if a.Uses == 0 {
   261  			if x.debug > 1 {
   262  				x.Printf("...marking %v unused\n", a.LongString())
   263  			}
   264  			x.invalidateRecursively(a)
   265  		}
   266  	}
   267  	v.Type = types.NewResults(append(abi.RegisterTypes(aux.abiInfo.OutParams()), types.TypeMem))
   268  	return
   269  }
   270  
   271  func (x *expandState) rewriteCallArgs(v *Value, firstArg int) {
   272  	if x.debug > 1 {
   273  		x.indent(3)
   274  		defer x.indent(-3)
   275  		x.Printf("rewriteCallArgs(%s; %d)\n", v.LongString(), firstArg)
   276  	}
   277  	// Thread the stores on the memory arg
   278  	aux := v.Aux.(*AuxCall)
   279  	m0 := v.MemoryArg()
   280  	mem := m0
   281  	allResults := []*Value{}
   282  	oldArgs := []*Value{}
   283  	argsWithoutMem := v.Args[firstArg : len(v.Args)-1] // Also strip closure/interface Op-specific args
   284  
   285  	sp := x.sp
   286  	if v.Op == OpTailLECall {
   287  		// For tail call, we unwind the frame before the call so we'll use the caller's
   288  		// SP.
   289  		sp = v.Block.NewValue1(src.NoXPos, OpGetCallerSP, x.typs.Uintptr, mem)
   290  	}
   291  
   292  	for i, a := range argsWithoutMem { // skip leading non-parameter SSA Args and trailing mem SSA Arg.
   293  		oldArgs = append(oldArgs, a)
   294  		auxI := int64(i)
   295  		aRegs := aux.RegsOfArg(auxI)
   296  		aType := aux.TypeOfArg(auxI)
   297  
   298  		if a.Op == OpDereference {
   299  			a.Op = OpLoad
   300  		}
   301  		var rc registerCursor
   302  		var result *[]*Value
   303  		var aOffset int64
   304  		if len(aRegs) > 0 {
   305  			result = &allResults
   306  		} else {
   307  			aOffset = aux.OffsetOfArg(auxI)
   308  		}
   309  		if v.Op == OpTailLECall && a.Op == OpArg && a.AuxInt == 0 {
   310  			// It's common for a tail call passing the same arguments (e.g. method wrapper),
   311  			// so this would be a self copy. Detect this and optimize it out.
   312  			n := a.Aux.(*ir.Name)
   313  			if n.Class == ir.PPARAM && n.FrameOffset()+x.f.Config.ctxt.Arch.FixedFrameSize == aOffset {
   314  				continue
   315  			}
   316  		}
   317  		if x.debug > 1 {
   318  			x.Printf("...storeArg %s, %v, %d\n", a.LongString(), aType, aOffset)
   319  		}
   320  
   321  		rc.init(aRegs, aux.abiInfo, result, sp, aOffset)
   322  		mem = x.decomposeAsNecessary(v.Pos, v.Block, a, mem, rc)
   323  	}
   324  	var preArgStore [2]*Value
   325  	preArgs := append(preArgStore[:0], v.Args[0:firstArg]...)
   326  	v.resetArgs()
   327  	v.AddArgs(preArgs...)
   328  	v.AddArgs(allResults...)
   329  	v.AddArg(mem)
   330  	for _, a := range oldArgs {
   331  		if a.Uses == 0 {
   332  			x.invalidateRecursively(a)
   333  		}
   334  	}
   335  
   336  	return
   337  }
   338  
   339  func (x *expandState) decomposePair(pos src.XPos, b *Block, a, mem *Value, t0, t1 *types.Type, o0, o1 Op, rc *registerCursor) *Value {
   340  	e := b.NewValue1(pos, o0, t0, a)
   341  	pos = pos.WithNotStmt()
   342  	mem = x.decomposeAsNecessary(pos, b, e, mem, rc.next(t0))
   343  	e = b.NewValue1(pos, o1, t1, a)
   344  	mem = x.decomposeAsNecessary(pos, b, e, mem, rc.next(t1))
   345  	return mem
   346  }
   347  
   348  func (x *expandState) decomposeOne(pos src.XPos, b *Block, a, mem *Value, t0 *types.Type, o0 Op, rc *registerCursor) *Value {
   349  	e := b.NewValue1(pos, o0, t0, a)
   350  	pos = pos.WithNotStmt()
   351  	mem = x.decomposeAsNecessary(pos, b, e, mem, rc.next(t0))
   352  	return mem
   353  }
   354  
   355  // decomposeAsNecessary converts a value (perhaps an aggregate) passed to a call or returned by a function,
   356  // into the appropriate sequence of stores and register assignments to transmit that value in a given ABI, and
   357  // returns the current memory after this convert/rewrite (it may be the input memory, perhaps stores were needed.)
   358  // 'pos' is the source position all this is tied to
   359  // 'b' is the enclosing block
   360  // 'a' is the value to decompose
   361  // 'm0' is the input memory arg used for the first store (or returned if there are no stores)
   362  // 'rc' is a registerCursor which identifies the register/memory destination for the value
   363  func (x *expandState) decomposeAsNecessary(pos src.XPos, b *Block, a, m0 *Value, rc registerCursor) *Value {
   364  	if x.debug > 1 {
   365  		x.indent(3)
   366  		defer x.indent(-3)
   367  	}
   368  	at := a.Type
   369  	if at.Size() == 0 {
   370  		return m0
   371  	}
   372  	if a.Op == OpDereference {
   373  		a.Op = OpLoad // For purposes of parameter passing expansion, a Dereference is a Load.
   374  	}
   375  
   376  	if !rc.hasRegs() && !CanSSA(at) {
   377  		dst := x.offsetFrom(b, rc.storeDest, rc.storeOffset, types.NewPtr(at))
   378  		if x.debug > 1 {
   379  			x.Printf("...recur store %s at %s\n", a.LongString(), dst.LongString())
   380  		}
   381  		if a.Op == OpLoad {
   382  			m0 = b.NewValue3A(pos, OpMove, types.TypeMem, at, dst, a.Args[0], m0)
   383  			m0.AuxInt = at.Size()
   384  			return m0
   385  		} else {
   386  			panic(fmt.Errorf("Store of not a load"))
   387  		}
   388  	}
   389  
   390  	mem := m0
   391  	switch at.Kind() {
   392  	case types.TARRAY:
   393  		et := at.Elem()
   394  		for i := int64(0); i < at.NumElem(); i++ {
   395  			e := b.NewValue1I(pos, OpArraySelect, et, i, a)
   396  			pos = pos.WithNotStmt()
   397  			mem = x.decomposeAsNecessary(pos, b, e, mem, rc.next(et))
   398  		}
   399  		return mem
   400  
   401  	case types.TSTRUCT:
   402  		for i := 0; i < at.NumFields(); i++ {
   403  			et := at.Field(i).Type // might need to read offsets from the fields
   404  			e := b.NewValue1I(pos, OpStructSelect, et, int64(i), a)
   405  			pos = pos.WithNotStmt()
   406  			if x.debug > 1 {
   407  				x.Printf("...recur decompose %s, %v\n", e.LongString(), et)
   408  			}
   409  			mem = x.decomposeAsNecessary(pos, b, e, mem, rc.next(et))
   410  		}
   411  		return mem
   412  
   413  	case types.TSLICE:
   414  		mem = x.decomposeOne(pos, b, a, mem, at.Elem().PtrTo(), OpSlicePtr, &rc)
   415  		pos = pos.WithNotStmt()
   416  		mem = x.decomposeOne(pos, b, a, mem, x.typs.Int, OpSliceLen, &rc)
   417  		return x.decomposeOne(pos, b, a, mem, x.typs.Int, OpSliceCap, &rc)
   418  
   419  	case types.TSTRING:
   420  		return x.decomposePair(pos, b, a, mem, x.typs.BytePtr, x.typs.Int, OpStringPtr, OpStringLen, &rc)
   421  
   422  	case types.TINTER:
   423  		mem = x.decomposeOne(pos, b, a, mem, x.typs.Uintptr, OpITab, &rc)
   424  		pos = pos.WithNotStmt()
   425  		// Immediate interfaces cause so many headaches.
   426  		if a.Op == OpIMake {
   427  			data := a.Args[1]
   428  			for data.Op == OpStructMake || data.Op == OpArrayMake1 {
   429  				data = data.Args[0]
   430  			}
   431  			return x.decomposeAsNecessary(pos, b, data, mem, rc.next(data.Type))
   432  		}
   433  		return x.decomposeOne(pos, b, a, mem, x.typs.BytePtr, OpIData, &rc)
   434  
   435  	case types.TCOMPLEX64:
   436  		return x.decomposePair(pos, b, a, mem, x.typs.Float32, x.typs.Float32, OpComplexReal, OpComplexImag, &rc)
   437  
   438  	case types.TCOMPLEX128:
   439  		return x.decomposePair(pos, b, a, mem, x.typs.Float64, x.typs.Float64, OpComplexReal, OpComplexImag, &rc)
   440  
   441  	case types.TINT64:
   442  		if at.Size() > x.regSize {
   443  			return x.decomposePair(pos, b, a, mem, x.firstType, x.secondType, x.firstOp, x.secondOp, &rc)
   444  		}
   445  	case types.TUINT64:
   446  		if at.Size() > x.regSize {
   447  			return x.decomposePair(pos, b, a, mem, x.typs.UInt32, x.typs.UInt32, x.firstOp, x.secondOp, &rc)
   448  		}
   449  	}
   450  
   451  	// An atomic type, either record the register or store it and update the memory.
   452  
   453  	if rc.hasRegs() {
   454  		if x.debug > 1 {
   455  			x.Printf("...recur addArg %s\n", a.LongString())
   456  		}
   457  		rc.addArg(a)
   458  	} else {
   459  		dst := x.offsetFrom(b, rc.storeDest, rc.storeOffset, types.NewPtr(at))
   460  		if x.debug > 1 {
   461  			x.Printf("...recur store %s at %s\n", a.LongString(), dst.LongString())
   462  		}
   463  		mem = b.NewValue3A(pos, OpStore, types.TypeMem, at, dst, a, mem)
   464  	}
   465  
   466  	return mem
   467  }
   468  
   469  // Convert scalar OpArg into the proper OpWhateverArg instruction
   470  // Convert scalar OpSelectN into perhaps-differently-indexed OpSelectN
   471  // Convert aggregate OpArg into Make of its parts (which are eventually scalars)
   472  // Convert aggregate OpSelectN into Make of its parts (which are eventually scalars)
   473  // Returns the converted value.
   474  //
   475  //   - "pos" the position for any generated instructions
   476  //   - "b" the block for any generated instructions
   477  //   - "container" the outermost OpArg/OpSelectN
   478  //   - "a" the instruction to overwrite, if any (only the outermost caller)
   479  //   - "m0" the memory arg for any loads that are necessary
   480  //   - "at" the type of the Arg/part
   481  //   - "rc" the register/memory cursor locating the various parts of the Arg.
   482  func (x *expandState) rewriteSelectOrArg(pos src.XPos, b *Block, container, a, m0 *Value, at *types.Type, rc registerCursor) *Value {
   483  
   484  	if at == types.TypeMem {
   485  		a.copyOf(m0)
   486  		return a
   487  	}
   488  
   489  	makeOf := func(a *Value, op Op, args []*Value) *Value {
   490  		if a == nil {
   491  			a = b.NewValue0(pos, op, at)
   492  			a.AddArgs(args...)
   493  		} else {
   494  			a.resetArgs()
   495  			a.Aux, a.AuxInt = nil, 0
   496  			a.Pos, a.Op, a.Type = pos, op, at
   497  			a.AddArgs(args...)
   498  		}
   499  		return a
   500  	}
   501  
   502  	if at.Size() == 0 {
   503  		// For consistency, create these values even though they'll ultimately be unused
   504  		if at.IsArray() {
   505  			return makeOf(a, OpArrayMake0, nil)
   506  		}
   507  		if at.IsStruct() {
   508  			return makeOf(a, OpStructMake, nil)
   509  		}
   510  		return a
   511  	}
   512  
   513  	sk := selKey{from: container, size: 0, offsetOrIndex: rc.storeOffset, typ: at}
   514  	dupe := x.commonSelectors[sk]
   515  	if dupe != nil {
   516  		if a == nil {
   517  			return dupe
   518  		}
   519  		a.copyOf(dupe)
   520  		return a
   521  	}
   522  
   523  	var argStore [10]*Value
   524  	args := argStore[:0]
   525  
   526  	addArg := func(a0 *Value) {
   527  		if a0 == nil {
   528  			as := "<nil>"
   529  			if a != nil {
   530  				as = a.LongString()
   531  			}
   532  			panic(fmt.Errorf("a0 should not be nil, a=%v, container=%v, at=%v", as, container.LongString(), at))
   533  		}
   534  		args = append(args, a0)
   535  	}
   536  
   537  	switch at.Kind() {
   538  	case types.TARRAY:
   539  		et := at.Elem()
   540  		for i := int64(0); i < at.NumElem(); i++ {
   541  			e := x.rewriteSelectOrArg(pos, b, container, nil, m0, et, rc.next(et))
   542  			addArg(e)
   543  		}
   544  		a = makeOf(a, OpArrayMake1, args)
   545  		x.commonSelectors[sk] = a
   546  		return a
   547  
   548  	case types.TSTRUCT:
   549  		// Assume ssagen/ssa.go (in buildssa) spills large aggregates so they won't appear here.
   550  		for i := 0; i < at.NumFields(); i++ {
   551  			et := at.Field(i).Type
   552  			e := x.rewriteSelectOrArg(pos, b, container, nil, m0, et, rc.next(et))
   553  			if e == nil {
   554  				panic(fmt.Errorf("nil e, et=%v, et.Size()=%d, i=%d", et, et.Size(), i))
   555  			}
   556  			addArg(e)
   557  			pos = pos.WithNotStmt()
   558  		}
   559  		if at.NumFields() > 4 {
   560  			panic(fmt.Errorf("Too many fields (%d, %d bytes), container=%s", at.NumFields(), at.Size(), container.LongString()))
   561  		}
   562  		a = makeOf(a, OpStructMake, args)
   563  		x.commonSelectors[sk] = a
   564  		return a
   565  
   566  	case types.TSLICE:
   567  		addArg(x.rewriteSelectOrArg(pos, b, container, nil, m0, at.Elem().PtrTo(), rc.next(x.typs.BytePtr)))
   568  		pos = pos.WithNotStmt()
   569  		addArg(x.rewriteSelectOrArg(pos, b, container, nil, m0, x.typs.Int, rc.next(x.typs.Int)))
   570  		addArg(x.rewriteSelectOrArg(pos, b, container, nil, m0, x.typs.Int, rc.next(x.typs.Int)))
   571  		a = makeOf(a, OpSliceMake, args)
   572  		x.commonSelectors[sk] = a
   573  		return a
   574  
   575  	case types.TSTRING:
   576  		addArg(x.rewriteSelectOrArg(pos, b, container, nil, m0, x.typs.BytePtr, rc.next(x.typs.BytePtr)))
   577  		pos = pos.WithNotStmt()
   578  		addArg(x.rewriteSelectOrArg(pos, b, container, nil, m0, x.typs.Int, rc.next(x.typs.Int)))
   579  		a = makeOf(a, OpStringMake, args)
   580  		x.commonSelectors[sk] = a
   581  		return a
   582  
   583  	case types.TINTER:
   584  		addArg(x.rewriteSelectOrArg(pos, b, container, nil, m0, x.typs.Uintptr, rc.next(x.typs.Uintptr)))
   585  		pos = pos.WithNotStmt()
   586  		addArg(x.rewriteSelectOrArg(pos, b, container, nil, m0, x.typs.BytePtr, rc.next(x.typs.BytePtr)))
   587  		a = makeOf(a, OpIMake, args)
   588  		x.commonSelectors[sk] = a
   589  		return a
   590  
   591  	case types.TCOMPLEX64:
   592  		addArg(x.rewriteSelectOrArg(pos, b, container, nil, m0, x.typs.Float32, rc.next(x.typs.Float32)))
   593  		pos = pos.WithNotStmt()
   594  		addArg(x.rewriteSelectOrArg(pos, b, container, nil, m0, x.typs.Float32, rc.next(x.typs.Float32)))
   595  		a = makeOf(a, OpComplexMake, args)
   596  		x.commonSelectors[sk] = a
   597  		return a
   598  
   599  	case types.TCOMPLEX128:
   600  		addArg(x.rewriteSelectOrArg(pos, b, container, nil, m0, x.typs.Float64, rc.next(x.typs.Float64)))
   601  		pos = pos.WithNotStmt()
   602  		addArg(x.rewriteSelectOrArg(pos, b, container, nil, m0, x.typs.Float64, rc.next(x.typs.Float64)))
   603  		a = makeOf(a, OpComplexMake, args)
   604  		x.commonSelectors[sk] = a
   605  		return a
   606  
   607  	case types.TINT64:
   608  		if at.Size() > x.regSize {
   609  			addArg(x.rewriteSelectOrArg(pos, b, container, nil, m0, x.firstType, rc.next(x.firstType)))
   610  			pos = pos.WithNotStmt()
   611  			addArg(x.rewriteSelectOrArg(pos, b, container, nil, m0, x.secondType, rc.next(x.secondType)))
   612  			if !x.f.Config.BigEndian {
   613  				// Int64Make args are big, little
   614  				args[0], args[1] = args[1], args[0]
   615  			}
   616  			a = makeOf(a, OpInt64Make, args)
   617  			x.commonSelectors[sk] = a
   618  			return a
   619  		}
   620  	case types.TUINT64:
   621  		if at.Size() > x.regSize {
   622  			addArg(x.rewriteSelectOrArg(pos, b, container, nil, m0, x.typs.UInt32, rc.next(x.typs.UInt32)))
   623  			pos = pos.WithNotStmt()
   624  			addArg(x.rewriteSelectOrArg(pos, b, container, nil, m0, x.typs.UInt32, rc.next(x.typs.UInt32)))
   625  			if !x.f.Config.BigEndian {
   626  				// Int64Make args are big, little
   627  				args[0], args[1] = args[1], args[0]
   628  			}
   629  			a = makeOf(a, OpInt64Make, args)
   630  			x.commonSelectors[sk] = a
   631  			return a
   632  		}
   633  	}
   634  
   635  	// An atomic type, either record the register or store it and update the memory.
   636  
   637  	// Depending on the container Op, the leaves are either OpSelectN or OpArg{Int,Float}Reg
   638  
   639  	if container.Op == OpArg {
   640  		if rc.hasRegs() {
   641  			op, i := rc.ArgOpAndRegisterFor()
   642  			name := container.Aux.(*ir.Name)
   643  			a = makeOf(a, op, nil)
   644  			a.AuxInt = i
   645  			a.Aux = &AuxNameOffset{name, rc.storeOffset}
   646  		} else {
   647  			key := selKey{container, rc.storeOffset, at.Size(), at}
   648  			w := x.commonArgs[key]
   649  			if w != nil && w.Uses != 0 {
   650  				if a == nil {
   651  					a = w
   652  				} else {
   653  					a.copyOf(w)
   654  				}
   655  			} else {
   656  				if a == nil {
   657  					aux := container.Aux
   658  					auxInt := container.AuxInt + rc.storeOffset
   659  					a = container.Block.NewValue0IA(container.Pos, OpArg, at, auxInt, aux)
   660  				} else {
   661  					// do nothing, the original should be okay.
   662  				}
   663  				x.commonArgs[key] = a
   664  			}
   665  		}
   666  	} else if container.Op == OpSelectN {
   667  		call := container.Args[0]
   668  		aux := call.Aux.(*AuxCall)
   669  		which := container.AuxInt
   670  
   671  		if at == types.TypeMem {
   672  			if a != m0 || a != x.memForCall[call.ID] {
   673  				panic(fmt.Errorf("Memories %s, %s, and %s should all be equal after %s", a.LongString(), m0.LongString(), x.memForCall[call.ID], call.LongString()))
   674  			}
   675  		} else if rc.hasRegs() {
   676  			firstReg := uint32(0)
   677  			for i := 0; i < int(which); i++ {
   678  				firstReg += uint32(len(aux.abiInfo.OutParam(i).Registers))
   679  			}
   680  			reg := int64(rc.nextSlice + Abi1RO(firstReg))
   681  			a = makeOf(a, OpSelectN, []*Value{call})
   682  			a.AuxInt = reg
   683  		} else {
   684  			off := x.offsetFrom(x.f.Entry, x.sp, rc.storeOffset+aux.OffsetOfResult(which), types.NewPtr(at))
   685  			a = makeOf(a, OpLoad, []*Value{off, m0})
   686  		}
   687  
   688  	} else {
   689  		panic(fmt.Errorf("Expected container OpArg or OpSelectN, saw %v instead", container.LongString()))
   690  	}
   691  
   692  	x.commonSelectors[sk] = a
   693  	return a
   694  }
   695  
   696  // rewriteWideSelectToStores handles the case of a SelectN'd result from a function call that is too large for SSA,
   697  // but is transferred in registers.  In this case the register cursor tracks both operands; the register sources and
   698  // the memory destinations.
   699  // This returns the memory flowing out of the last store
   700  func (x *expandState) rewriteWideSelectToStores(pos src.XPos, b *Block, container, m0 *Value, at *types.Type, rc registerCursor) *Value {
   701  
   702  	if at.Size() == 0 {
   703  		return m0
   704  	}
   705  
   706  	switch at.Kind() {
   707  	case types.TARRAY:
   708  		et := at.Elem()
   709  		for i := int64(0); i < at.NumElem(); i++ {
   710  			m0 = x.rewriteWideSelectToStores(pos, b, container, m0, et, rc.next(et))
   711  		}
   712  		return m0
   713  
   714  	case types.TSTRUCT:
   715  		// Assume ssagen/ssa.go (in buildssa) spills large aggregates so they won't appear here.
   716  		for i := 0; i < at.NumFields(); i++ {
   717  			et := at.Field(i).Type
   718  			m0 = x.rewriteWideSelectToStores(pos, b, container, m0, et, rc.next(et))
   719  			pos = pos.WithNotStmt()
   720  		}
   721  		return m0
   722  
   723  	case types.TSLICE:
   724  		m0 = x.rewriteWideSelectToStores(pos, b, container, m0, at.Elem().PtrTo(), rc.next(x.typs.BytePtr))
   725  		pos = pos.WithNotStmt()
   726  		m0 = x.rewriteWideSelectToStores(pos, b, container, m0, x.typs.Int, rc.next(x.typs.Int))
   727  		m0 = x.rewriteWideSelectToStores(pos, b, container, m0, x.typs.Int, rc.next(x.typs.Int))
   728  		return m0
   729  
   730  	case types.TSTRING:
   731  		m0 = x.rewriteWideSelectToStores(pos, b, container, m0, x.typs.BytePtr, rc.next(x.typs.BytePtr))
   732  		pos = pos.WithNotStmt()
   733  		m0 = x.rewriteWideSelectToStores(pos, b, container, m0, x.typs.Int, rc.next(x.typs.Int))
   734  		return m0
   735  
   736  	case types.TINTER:
   737  		m0 = x.rewriteWideSelectToStores(pos, b, container, m0, x.typs.Uintptr, rc.next(x.typs.Uintptr))
   738  		pos = pos.WithNotStmt()
   739  		m0 = x.rewriteWideSelectToStores(pos, b, container, m0, x.typs.BytePtr, rc.next(x.typs.BytePtr))
   740  		return m0
   741  
   742  	case types.TCOMPLEX64:
   743  		m0 = x.rewriteWideSelectToStores(pos, b, container, m0, x.typs.Float32, rc.next(x.typs.Float32))
   744  		pos = pos.WithNotStmt()
   745  		m0 = x.rewriteWideSelectToStores(pos, b, container, m0, x.typs.Float32, rc.next(x.typs.Float32))
   746  		return m0
   747  
   748  	case types.TCOMPLEX128:
   749  		m0 = x.rewriteWideSelectToStores(pos, b, container, m0, x.typs.Float64, rc.next(x.typs.Float64))
   750  		pos = pos.WithNotStmt()
   751  		m0 = x.rewriteWideSelectToStores(pos, b, container, m0, x.typs.Float64, rc.next(x.typs.Float64))
   752  		return m0
   753  
   754  	case types.TINT64:
   755  		if at.Size() > x.regSize {
   756  			m0 = x.rewriteWideSelectToStores(pos, b, container, m0, x.firstType, rc.next(x.firstType))
   757  			pos = pos.WithNotStmt()
   758  			m0 = x.rewriteWideSelectToStores(pos, b, container, m0, x.secondType, rc.next(x.secondType))
   759  			return m0
   760  		}
   761  	case types.TUINT64:
   762  		if at.Size() > x.regSize {
   763  			m0 = x.rewriteWideSelectToStores(pos, b, container, m0, x.typs.UInt32, rc.next(x.typs.UInt32))
   764  			pos = pos.WithNotStmt()
   765  			m0 = x.rewriteWideSelectToStores(pos, b, container, m0, x.typs.UInt32, rc.next(x.typs.UInt32))
   766  			return m0
   767  		}
   768  	}
   769  
   770  	// TODO could change treatment of too-large OpArg, would deal with it here.
   771  	if container.Op == OpSelectN {
   772  		call := container.Args[0]
   773  		aux := call.Aux.(*AuxCall)
   774  		which := container.AuxInt
   775  
   776  		if rc.hasRegs() {
   777  			firstReg := uint32(0)
   778  			for i := 0; i < int(which); i++ {
   779  				firstReg += uint32(len(aux.abiInfo.OutParam(i).Registers))
   780  			}
   781  			reg := int64(rc.nextSlice + Abi1RO(firstReg))
   782  			a := b.NewValue1I(pos, OpSelectN, at, reg, call)
   783  			dst := x.offsetFrom(b, rc.storeDest, rc.storeOffset, types.NewPtr(at))
   784  			m0 = b.NewValue3A(pos, OpStore, types.TypeMem, at, dst, a, m0)
   785  		} else {
   786  			panic(fmt.Errorf("Expected rc to have registers"))
   787  		}
   788  	} else {
   789  		panic(fmt.Errorf("Expected container OpSelectN, saw %v instead", container.LongString()))
   790  	}
   791  	return m0
   792  }
   793  
   794  func isBlockMultiValueExit(b *Block) bool {
   795  	return (b.Kind == BlockRet || b.Kind == BlockRetJmp) && b.Controls[0] != nil && b.Controls[0].Op == OpMakeResult
   796  }
   797  
   798  type Abi1RO uint8 // An offset within a parameter's slice of register indices, for abi1.
   799  
   800  // A registerCursor tracks which register is used for an Arg or regValues, or a piece of such.
   801  type registerCursor struct {
   802  	storeDest   *Value // if there are no register targets, then this is the base of the store.
   803  	storeOffset int64
   804  	regs        []abi.RegIndex // the registers available for this Arg/result (which is all in registers or not at all)
   805  	nextSlice   Abi1RO         // the next register/register-slice offset
   806  	config      *abi.ABIConfig
   807  	regValues   *[]*Value // values assigned to registers accumulate here
   808  }
   809  
   810  func (c *registerCursor) String() string {
   811  	dest := "<none>"
   812  	if c.storeDest != nil {
   813  		dest = fmt.Sprintf("%s+%d", c.storeDest.String(), c.storeOffset)
   814  	}
   815  	regs := "<none>"
   816  	if c.regValues != nil {
   817  		regs = ""
   818  		for i, x := range *c.regValues {
   819  			if i > 0 {
   820  				regs = regs + "; "
   821  			}
   822  			regs = regs + x.LongString()
   823  		}
   824  	}
   825  
   826  	// not printing the config because that has not been useful
   827  	return fmt.Sprintf("RCSR{storeDest=%v, regsLen=%d, nextSlice=%d, regValues=[%s]}", dest, len(c.regs), c.nextSlice, regs)
   828  }
   829  
   830  // next effectively post-increments the register cursor; the receiver is advanced,
   831  // the (aligned) old value is returned.
   832  func (c *registerCursor) next(t *types.Type) registerCursor {
   833  	c.storeOffset = types.RoundUp(c.storeOffset, t.Alignment())
   834  	rc := *c
   835  	c.storeOffset = types.RoundUp(c.storeOffset+t.Size(), t.Alignment())
   836  	if int(c.nextSlice) < len(c.regs) {
   837  		w := c.config.NumParamRegs(t)
   838  		c.nextSlice += Abi1RO(w)
   839  	}
   840  	return rc
   841  }
   842  
   843  // plus returns a register cursor offset from the original, without modifying the original.
   844  func (c *registerCursor) plus(regWidth Abi1RO) registerCursor {
   845  	rc := *c
   846  	rc.nextSlice += regWidth
   847  	return rc
   848  }
   849  
   850  // at returns the register cursor for component i of t, where the first
   851  // component is numbered 0.
   852  func (c *registerCursor) at(t *types.Type, i int) registerCursor {
   853  	rc := *c
   854  	if i == 0 || len(c.regs) == 0 {
   855  		return rc
   856  	}
   857  	if t.IsArray() {
   858  		w := c.config.NumParamRegs(t.Elem())
   859  		rc.nextSlice += Abi1RO(i * w)
   860  		return rc
   861  	}
   862  	if t.IsStruct() {
   863  		for j := 0; j < i; j++ {
   864  			rc.next(t.FieldType(j))
   865  		}
   866  		return rc
   867  	}
   868  	panic("Haven't implemented this case yet, do I need to?")
   869  }
   870  
   871  func (c *registerCursor) init(regs []abi.RegIndex, info *abi.ABIParamResultInfo, result *[]*Value, storeDest *Value, storeOffset int64) {
   872  	c.regs = regs
   873  	c.nextSlice = 0
   874  	c.storeOffset = storeOffset
   875  	c.storeDest = storeDest
   876  	c.config = info.Config()
   877  	c.regValues = result
   878  }
   879  
   880  func (c *registerCursor) addArg(v *Value) {
   881  	*c.regValues = append(*c.regValues, v)
   882  }
   883  
   884  func (c *registerCursor) hasRegs() bool {
   885  	return len(c.regs) > 0
   886  }
   887  
   888  func (c *registerCursor) ArgOpAndRegisterFor() (Op, int64) {
   889  	r := c.regs[c.nextSlice]
   890  	return ArgOpAndRegisterFor(r, c.config)
   891  }
   892  
   893  // ArgOpAndRegisterFor converts an abi register index into an ssa Op and corresponding
   894  // arg register index.
   895  func ArgOpAndRegisterFor(r abi.RegIndex, abiConfig *abi.ABIConfig) (Op, int64) {
   896  	i := abiConfig.FloatIndexFor(r)
   897  	if i >= 0 { // float PR
   898  		return OpArgFloatReg, i
   899  	}
   900  	return OpArgIntReg, int64(r)
   901  }
   902  
   903  type selKey struct {
   904  	from          *Value // what is selected from
   905  	offsetOrIndex int64  // whatever is appropriate for the selector
   906  	size          int64
   907  	typ           *types.Type
   908  }
   909  
   910  type expandState struct {
   911  	f       *Func
   912  	debug   int // odd values log lost statement markers, so likely settings are 1 (stmts), 2 (expansion), and 3 (both)
   913  	regSize int64
   914  	sp      *Value
   915  	typs    *Types
   916  
   917  	firstOp    Op          // for 64-bit integers on 32-bit machines, first word in memory
   918  	secondOp   Op          // for 64-bit integers on 32-bit machines, second word in memory
   919  	firstType  *types.Type // first half type, for Int64
   920  	secondType *types.Type // second half type, for Int64
   921  
   922  	wideSelects     map[*Value]*Value // Selects that are not SSA-able, mapped to consuming stores.
   923  	commonSelectors map[selKey]*Value // used to de-dupe selectors
   924  	commonArgs      map[selKey]*Value // used to de-dupe OpArg/OpArgIntReg/OpArgFloatReg
   925  	memForCall      map[ID]*Value     // For a call, need to know the unique selector that gets the mem.
   926  	indentLevel     int               // Indentation for debugging recursion
   927  }
   928  
   929  // intPairTypes returns the pair of 32-bit int types needed to encode a 64-bit integer type on a target
   930  // that has no 64-bit integer registers.
   931  func (x *expandState) intPairTypes(et types.Kind) (tHi, tLo *types.Type) {
   932  	tHi = x.typs.UInt32
   933  	if et == types.TINT64 {
   934  		tHi = x.typs.Int32
   935  	}
   936  	tLo = x.typs.UInt32
   937  	return
   938  }
   939  
   940  // offsetFrom creates an offset from a pointer, simplifying chained offsets and offsets from SP
   941  func (x *expandState) offsetFrom(b *Block, from *Value, offset int64, pt *types.Type) *Value {
   942  	ft := from.Type
   943  	if offset == 0 {
   944  		if ft == pt {
   945  			return from
   946  		}
   947  		// This captures common, (apparently) safe cases.  The unsafe cases involve ft == uintptr
   948  		if (ft.IsPtr() || ft.IsUnsafePtr()) && pt.IsPtr() {
   949  			return from
   950  		}
   951  	}
   952  	// Simplify, canonicalize
   953  	for from.Op == OpOffPtr {
   954  		offset += from.AuxInt
   955  		from = from.Args[0]
   956  	}
   957  	if from == x.sp {
   958  		return x.f.ConstOffPtrSP(pt, offset, x.sp)
   959  	}
   960  	return b.NewValue1I(from.Pos.WithNotStmt(), OpOffPtr, pt, offset, from)
   961  }
   962  
   963  func (x *expandState) regWidth(t *types.Type) Abi1RO {
   964  	return Abi1RO(x.f.ABI1.NumParamRegs(t))
   965  }
   966  
   967  // regOffset returns the register offset of the i'th element of type t
   968  func (x *expandState) regOffset(t *types.Type, i int) Abi1RO {
   969  	// TODO maybe cache this in a map if profiling recommends.
   970  	if i == 0 {
   971  		return 0
   972  	}
   973  	if t.IsArray() {
   974  		return Abi1RO(i) * x.regWidth(t.Elem())
   975  	}
   976  	if t.IsStruct() {
   977  		k := Abi1RO(0)
   978  		for j := 0; j < i; j++ {
   979  			k += x.regWidth(t.FieldType(j))
   980  		}
   981  		return k
   982  	}
   983  	panic("Haven't implemented this case yet, do I need to?")
   984  }
   985  
   986  // prAssignForArg returns the ABIParamAssignment for v, assumed to be an OpArg.
   987  func (x *expandState) prAssignForArg(v *Value) *abi.ABIParamAssignment {
   988  	if v.Op != OpArg {
   989  		panic(fmt.Errorf("Wanted OpArg, instead saw %s", v.LongString()))
   990  	}
   991  	return ParamAssignmentForArgName(x.f, v.Aux.(*ir.Name))
   992  }
   993  
   994  // ParamAssignmentForArgName returns the ABIParamAssignment for f's arg with matching name.
   995  func ParamAssignmentForArgName(f *Func, name *ir.Name) *abi.ABIParamAssignment {
   996  	abiInfo := f.OwnAux.abiInfo
   997  	ip := abiInfo.InParams()
   998  	for i, a := range ip {
   999  		if a.Name == name {
  1000  			return &ip[i]
  1001  		}
  1002  	}
  1003  	panic(fmt.Errorf("Did not match param %v in prInfo %+v", name, abiInfo.InParams()))
  1004  }
  1005  
  1006  // indent increments (or decrements) the indentation.
  1007  func (x *expandState) indent(n int) {
  1008  	x.indentLevel += n
  1009  }
  1010  
  1011  // Printf does an indented fmt.Printf on the format and args.
  1012  func (x *expandState) Printf(format string, a ...interface{}) (n int, err error) {
  1013  	if x.indentLevel > 0 {
  1014  		fmt.Printf("%[1]*s", x.indentLevel, "")
  1015  	}
  1016  	return fmt.Printf(format, a...)
  1017  }
  1018  
  1019  func (x *expandState) invalidateRecursively(a *Value) {
  1020  	var s string
  1021  	if x.debug > 0 {
  1022  		plus := " "
  1023  		if a.Pos.IsStmt() == src.PosIsStmt {
  1024  			plus = " +"
  1025  		}
  1026  		s = a.String() + plus + a.Pos.LineNumber() + " " + a.LongString()
  1027  		if x.debug > 1 {
  1028  			x.Printf("...marking %v unused\n", s)
  1029  		}
  1030  	}
  1031  	lost := a.invalidateRecursively()
  1032  	if x.debug&1 != 0 && lost { // For odd values of x.debug, do this.
  1033  		x.Printf("Lost statement marker in %s on former %s\n", base.Ctxt.Pkgpath+"."+x.f.Name, s)
  1034  	}
  1035  }
  1036  

View as plain text