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, OpTailLECallInter:
    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 OpTailLECallInter:
   203  			rewriteCall(v, OpTailCallInter, 1)
   204  		case OpClosureLECall:
   205  			rewriteCall(v, OpClosureCall, 2)
   206  		case OpInterLECall:
   207  			rewriteCall(v, OpInterCall, 1)
   208  		}
   209  	}
   210  
   211  	// Rewrite results from exit blocks
   212  	for _, b := range exitBlocks {
   213  		v := b.Controls[0]
   214  		x.rewriteFuncResults(v, b, f.OwnAux)
   215  		b.SetControl(v)
   216  	}
   217  
   218  }
   219  
   220  func (x *expandState) rewriteFuncResults(v *Value, b *Block, aux *AuxCall) {
   221  	// This is very similar to rewriteCallArgs
   222  	// differences:
   223  	// firstArg + preArgs
   224  	// sp vs auxBase
   225  
   226  	m0 := v.MemoryArg()
   227  	mem := m0
   228  
   229  	allResults := []*Value{}
   230  	var oldArgs []*Value
   231  	argsWithoutMem := v.Args[:len(v.Args)-1]
   232  
   233  	for j, a := range argsWithoutMem {
   234  		oldArgs = append(oldArgs, a)
   235  		i := int64(j)
   236  		auxType := aux.TypeOfResult(i)
   237  		auxBase := b.NewValue2A(v.Pos, OpLocalAddr, types.NewPtr(auxType), aux.NameOfResult(i), x.sp, mem)
   238  		auxOffset := int64(0)
   239  		aRegs := aux.RegsOfResult(int64(j))
   240  		if a.Op == OpDereference {
   241  			a.Op = OpLoad
   242  		}
   243  		var rc registerCursor
   244  		var result *[]*Value
   245  		if len(aRegs) > 0 {
   246  			result = &allResults
   247  		} else {
   248  			if a.Op == OpLoad && a.Args[0].Op == OpLocalAddr && a.Args[0].Aux == aux.NameOfResult(i) {
   249  				continue // Self move to output parameter
   250  			}
   251  		}
   252  		rc.init(aRegs, aux.abiInfo, result, auxBase, auxOffset)
   253  		mem = x.decomposeAsNecessary(v.Pos, b, a, mem, rc)
   254  	}
   255  	v.resetArgs()
   256  	v.AddArgs(allResults...)
   257  	v.AddArg(mem)
   258  	for _, a := range oldArgs {
   259  		if a.Uses == 0 {
   260  			if x.debug > 1 {
   261  				x.Printf("...marking %v unused\n", a.LongString())
   262  			}
   263  			x.invalidateRecursively(a)
   264  		}
   265  	}
   266  	v.Type = types.NewResults(append(abi.RegisterTypes(aux.abiInfo.OutParams()), types.TypeMem))
   267  	return
   268  }
   269  
   270  func (x *expandState) rewriteCallArgs(v *Value, firstArg int) {
   271  	if x.debug > 1 {
   272  		x.indent(3)
   273  		defer x.indent(-3)
   274  		x.Printf("rewriteCallArgs(%s; %d)\n", v.LongString(), firstArg)
   275  	}
   276  	// Thread the stores on the memory arg
   277  	aux := v.Aux.(*AuxCall)
   278  	m0 := v.MemoryArg()
   279  	mem := m0
   280  	allResults := []*Value{}
   281  	oldArgs := []*Value{}
   282  	argsWithoutMem := v.Args[firstArg : len(v.Args)-1] // Also strip closure/interface Op-specific args
   283  
   284  	sp := x.sp
   285  	if v.Op == OpTailLECall || v.Op == OpTailLECallInter {
   286  		// For tail call, we unwind the frame before the call so we'll use the caller's
   287  		// SP.
   288  		sp = v.Block.NewValue1(src.NoXPos, OpGetCallerSP, x.typs.Uintptr, mem)
   289  	}
   290  
   291  	for i, a := range argsWithoutMem { // skip leading non-parameter SSA Args and trailing mem SSA Arg.
   292  		oldArgs = append(oldArgs, a)
   293  		auxI := int64(i)
   294  		aRegs := aux.RegsOfArg(auxI)
   295  		aType := aux.TypeOfArg(auxI)
   296  
   297  		if a.Op == OpDereference {
   298  			a.Op = OpLoad
   299  		}
   300  		var rc registerCursor
   301  		var result *[]*Value
   302  		var aOffset int64
   303  		if len(aRegs) > 0 {
   304  			result = &allResults
   305  		} else {
   306  			aOffset = aux.OffsetOfArg(auxI)
   307  		}
   308  		if v.Op == OpTailLECall && a.Op == OpArg && a.AuxInt == 0 {
   309  			// It's common for a tail call passing the same arguments (e.g. method wrapper),
   310  			// so this would be a self copy. Detect this and optimize it out.
   311  			n := a.Aux.(*ir.Name)
   312  			if n.Class == ir.PPARAM && n.FrameOffset()+x.f.Config.ctxt.Arch.FixedFrameSize == aOffset {
   313  				continue
   314  			}
   315  		}
   316  		if x.debug > 1 {
   317  			x.Printf("...storeArg %s, %v, %d\n", a.LongString(), aType, aOffset)
   318  		}
   319  
   320  		rc.init(aRegs, aux.abiInfo, result, sp, aOffset)
   321  		mem = x.decomposeAsNecessary(v.Pos, v.Block, a, mem, rc)
   322  	}
   323  	var preArgStore [2]*Value
   324  	preArgs := append(preArgStore[:0], v.Args[0:firstArg]...)
   325  	v.resetArgs()
   326  	v.AddArgs(preArgs...)
   327  	v.AddArgs(allResults...)
   328  	v.AddArg(mem)
   329  	for _, a := range oldArgs {
   330  		if a.Uses == 0 {
   331  			x.invalidateRecursively(a)
   332  		}
   333  	}
   334  
   335  	return
   336  }
   337  
   338  func (x *expandState) decomposePair(pos src.XPos, b *Block, a, mem *Value, t0, t1 *types.Type, o0, o1 Op, rc *registerCursor) *Value {
   339  	e := b.NewValue1(pos, o0, t0, a)
   340  	pos = pos.WithNotStmt()
   341  	mem = x.decomposeAsNecessary(pos, b, e, mem, rc.next(t0))
   342  	e = b.NewValue1(pos, o1, t1, a)
   343  	mem = x.decomposeAsNecessary(pos, b, e, mem, rc.next(t1))
   344  	return mem
   345  }
   346  
   347  func (x *expandState) decomposeOne(pos src.XPos, b *Block, a, mem *Value, t0 *types.Type, o0 Op, rc *registerCursor) *Value {
   348  	e := b.NewValue1(pos, o0, t0, a)
   349  	pos = pos.WithNotStmt()
   350  	mem = x.decomposeAsNecessary(pos, b, e, mem, rc.next(t0))
   351  	return mem
   352  }
   353  
   354  // decomposeAsNecessary converts a value (perhaps an aggregate) passed to a call or returned by a function,
   355  // into the appropriate sequence of stores and register assignments to transmit that value in a given ABI, and
   356  // returns the current memory after this convert/rewrite (it may be the input memory, perhaps stores were needed.)
   357  // 'pos' is the source position all this is tied to
   358  // 'b' is the enclosing block
   359  // 'a' is the value to decompose
   360  // 'm0' is the input memory arg used for the first store (or returned if there are no stores)
   361  // 'rc' is a registerCursor which identifies the register/memory destination for the value
   362  func (x *expandState) decomposeAsNecessary(pos src.XPos, b *Block, a, m0 *Value, rc registerCursor) *Value {
   363  	if x.debug > 1 {
   364  		x.indent(3)
   365  		defer x.indent(-3)
   366  	}
   367  	at := a.Type
   368  	if at.Size() == 0 {
   369  		return m0
   370  	}
   371  	if a.Op == OpDereference {
   372  		a.Op = OpLoad // For purposes of parameter passing expansion, a Dereference is a Load.
   373  	}
   374  
   375  	if !rc.hasRegs() && !CanSSA(at) {
   376  		dst := x.offsetFrom(b, rc.storeDest, rc.storeOffset, types.NewPtr(at))
   377  		if x.debug > 1 {
   378  			x.Printf("...recur store %s at %s\n", a.LongString(), dst.LongString())
   379  		}
   380  		if a.Op == OpLoad {
   381  			m0 = b.NewValue3A(pos, OpMove, types.TypeMem, at, dst, a.Args[0], m0)
   382  			m0.AuxInt = at.Size()
   383  			return m0
   384  		} else {
   385  			panic(fmt.Errorf("Store of not a load"))
   386  		}
   387  	}
   388  
   389  	mem := m0
   390  	switch at.Kind() {
   391  	case types.TARRAY:
   392  		et := at.Elem()
   393  		for i := int64(0); i < at.NumElem(); i++ {
   394  			e := b.NewValue1I(pos, OpArraySelect, et, i, a)
   395  			pos = pos.WithNotStmt()
   396  			mem = x.decomposeAsNecessary(pos, b, e, mem, rc.next(et))
   397  		}
   398  		return mem
   399  
   400  	case types.TSTRUCT:
   401  		if at.IsSIMD() {
   402  			break // XXX
   403  		}
   404  		for i := 0; i < at.NumFields(); i++ {
   405  			et := at.Field(i).Type // might need to read offsets from the fields
   406  			e := b.NewValue1I(pos, OpStructSelect, et, int64(i), a)
   407  			pos = pos.WithNotStmt()
   408  			if x.debug > 1 {
   409  				x.Printf("...recur decompose %s, %v\n", e.LongString(), et)
   410  			}
   411  			mem = x.decomposeAsNecessary(pos, b, e, mem, rc.next(et))
   412  		}
   413  		return mem
   414  
   415  	case types.TSLICE:
   416  		mem = x.decomposeOne(pos, b, a, mem, at.Elem().PtrTo(), OpSlicePtr, &rc)
   417  		pos = pos.WithNotStmt()
   418  		mem = x.decomposeOne(pos, b, a, mem, x.typs.Int, OpSliceLen, &rc)
   419  		return x.decomposeOne(pos, b, a, mem, x.typs.Int, OpSliceCap, &rc)
   420  
   421  	case types.TSTRING:
   422  		return x.decomposePair(pos, b, a, mem, x.typs.BytePtr, x.typs.Int, OpStringPtr, OpStringLen, &rc)
   423  
   424  	case types.TINTER:
   425  		mem = x.decomposeOne(pos, b, a, mem, x.typs.Uintptr, OpITab, &rc)
   426  		pos = pos.WithNotStmt()
   427  		// Immediate interfaces cause so many headaches.
   428  		if a.Op == OpIMake {
   429  			data := a.Args[1]
   430  			for data.Op == OpStructMake || data.Op == OpArrayMake1 {
   431  				// A struct make might have a few zero-sized fields.
   432  				// Use the pointer-y one we know is there.
   433  				for _, a := range data.Args {
   434  					if a.Type.Size() > 0 {
   435  						data = a
   436  						break
   437  					}
   438  				}
   439  			}
   440  			return x.decomposeAsNecessary(pos, b, data, mem, rc.next(data.Type))
   441  		}
   442  		return x.decomposeOne(pos, b, a, mem, x.typs.BytePtr, OpIData, &rc)
   443  
   444  	case types.TCOMPLEX64:
   445  		return x.decomposePair(pos, b, a, mem, x.typs.Float32, x.typs.Float32, OpComplexReal, OpComplexImag, &rc)
   446  
   447  	case types.TCOMPLEX128:
   448  		return x.decomposePair(pos, b, a, mem, x.typs.Float64, x.typs.Float64, OpComplexReal, OpComplexImag, &rc)
   449  
   450  	case types.TINT64:
   451  		if at.Size() > x.regSize {
   452  			return x.decomposePair(pos, b, a, mem, x.firstType, x.secondType, x.firstOp, x.secondOp, &rc)
   453  		}
   454  	case types.TUINT64:
   455  		if at.Size() > x.regSize {
   456  			return x.decomposePair(pos, b, a, mem, x.typs.UInt32, x.typs.UInt32, x.firstOp, x.secondOp, &rc)
   457  		}
   458  	}
   459  
   460  	// An atomic type, either record the register or store it and update the memory.
   461  
   462  	if rc.hasRegs() {
   463  		if x.debug > 1 {
   464  			x.Printf("...recur addArg %s\n", a.LongString())
   465  		}
   466  		rc.addArg(a)
   467  	} else {
   468  		dst := x.offsetFrom(b, rc.storeDest, rc.storeOffset, types.NewPtr(at))
   469  		if x.debug > 1 {
   470  			x.Printf("...recur store %s at %s\n", a.LongString(), dst.LongString())
   471  		}
   472  		mem = b.NewValue3A(pos, OpStore, types.TypeMem, at, dst, a, mem)
   473  	}
   474  
   475  	return mem
   476  }
   477  
   478  // Convert scalar OpArg into the proper OpWhateverArg instruction
   479  // Convert scalar OpSelectN into perhaps-differently-indexed OpSelectN
   480  // Convert aggregate OpArg into Make of its parts (which are eventually scalars)
   481  // Convert aggregate OpSelectN into Make of its parts (which are eventually scalars)
   482  // Returns the converted value.
   483  //
   484  //   - "pos" the position for any generated instructions
   485  //   - "b" the block for any generated instructions
   486  //   - "container" the outermost OpArg/OpSelectN
   487  //   - "a" the instruction to overwrite, if any (only the outermost caller)
   488  //   - "m0" the memory arg for any loads that are necessary
   489  //   - "at" the type of the Arg/part
   490  //   - "rc" the register/memory cursor locating the various parts of the Arg.
   491  func (x *expandState) rewriteSelectOrArg(pos src.XPos, b *Block, container, a, m0 *Value, at *types.Type, rc registerCursor) *Value {
   492  
   493  	if at == types.TypeMem {
   494  		a.copyOf(m0)
   495  		return a
   496  	}
   497  
   498  	makeOf := func(a *Value, op Op, args []*Value) *Value {
   499  		if a == nil {
   500  			a = b.NewValue0(pos, op, at)
   501  			a.AddArgs(args...)
   502  		} else {
   503  			a.resetArgs()
   504  			a.Aux, a.AuxInt = nil, 0
   505  			a.Pos, a.Op, a.Type = pos, op, at
   506  			a.AddArgs(args...)
   507  		}
   508  		return a
   509  	}
   510  
   511  	if at.Size() == 0 {
   512  		// For consistency, create these values even though they'll ultimately be unused
   513  		return makeOf(a, OpEmpty, nil)
   514  	}
   515  
   516  	sk := selKey{from: container, size: 0, offsetOrIndex: rc.storeOffset, typ: at}
   517  	dupe := x.commonSelectors[sk]
   518  	if dupe != nil {
   519  		if a == nil {
   520  			return dupe
   521  		}
   522  		a.copyOf(dupe)
   523  		return a
   524  	}
   525  
   526  	var argStore [10]*Value
   527  	args := argStore[:0]
   528  
   529  	addArg := func(a0 *Value) {
   530  		if a0 == nil {
   531  			as := "<nil>"
   532  			if a != nil {
   533  				as = a.LongString()
   534  			}
   535  			panic(fmt.Errorf("a0 should not be nil, a=%v, container=%v, at=%v", as, container.LongString(), at))
   536  		}
   537  		args = append(args, a0)
   538  	}
   539  
   540  	switch at.Kind() {
   541  	case types.TARRAY:
   542  		et := at.Elem()
   543  		for i := int64(0); i < at.NumElem(); i++ {
   544  			e := x.rewriteSelectOrArg(pos, b, container, nil, m0, et, rc.next(et))
   545  			addArg(e)
   546  		}
   547  		a = makeOf(a, OpArrayMake1, args)
   548  		x.commonSelectors[sk] = a
   549  		return a
   550  
   551  	case types.TSTRUCT:
   552  		// Assume ssagen/ssa.go (in buildssa) spills large aggregates so they won't appear here.
   553  		if at.IsSIMD() {
   554  			break // XXX
   555  		}
   556  		for i := 0; i < at.NumFields(); i++ {
   557  			et := at.Field(i).Type
   558  			e := x.rewriteSelectOrArg(pos, b, container, nil, m0, et, rc.next(et))
   559  			if e == nil {
   560  				panic(fmt.Errorf("nil e, et=%v, et.Size()=%d, i=%d", et, et.Size(), i))
   561  			}
   562  			addArg(e)
   563  			pos = pos.WithNotStmt()
   564  		}
   565  		if at.NumFields() > MaxStruct && !types.IsDirectIface(at) {
   566  			panic(fmt.Errorf("Too many fields (%d, %d bytes), container=%s", at.NumFields(), at.Size(), container.LongString()))
   567  		}
   568  		a = makeOf(a, OpStructMake, args)
   569  		x.commonSelectors[sk] = a
   570  		return a
   571  
   572  	case types.TSLICE:
   573  		addArg(x.rewriteSelectOrArg(pos, b, container, nil, m0, at.Elem().PtrTo(), rc.next(x.typs.BytePtr)))
   574  		pos = pos.WithNotStmt()
   575  		addArg(x.rewriteSelectOrArg(pos, b, container, nil, m0, x.typs.Int, rc.next(x.typs.Int)))
   576  		addArg(x.rewriteSelectOrArg(pos, b, container, nil, m0, x.typs.Int, rc.next(x.typs.Int)))
   577  		a = makeOf(a, OpSliceMake, args)
   578  		x.commonSelectors[sk] = a
   579  		return a
   580  
   581  	case types.TSTRING:
   582  		addArg(x.rewriteSelectOrArg(pos, b, container, nil, m0, x.typs.BytePtr, rc.next(x.typs.BytePtr)))
   583  		pos = pos.WithNotStmt()
   584  		addArg(x.rewriteSelectOrArg(pos, b, container, nil, m0, x.typs.Int, rc.next(x.typs.Int)))
   585  		a = makeOf(a, OpStringMake, args)
   586  		x.commonSelectors[sk] = a
   587  		return a
   588  
   589  	case types.TINTER:
   590  		addArg(x.rewriteSelectOrArg(pos, b, container, nil, m0, x.typs.Uintptr, rc.next(x.typs.Uintptr)))
   591  		pos = pos.WithNotStmt()
   592  		addArg(x.rewriteSelectOrArg(pos, b, container, nil, m0, x.typs.BytePtr, rc.next(x.typs.BytePtr)))
   593  		a = makeOf(a, OpIMake, args)
   594  		x.commonSelectors[sk] = a
   595  		return a
   596  
   597  	case types.TCOMPLEX64:
   598  		addArg(x.rewriteSelectOrArg(pos, b, container, nil, m0, x.typs.Float32, rc.next(x.typs.Float32)))
   599  		pos = pos.WithNotStmt()
   600  		addArg(x.rewriteSelectOrArg(pos, b, container, nil, m0, x.typs.Float32, rc.next(x.typs.Float32)))
   601  		a = makeOf(a, OpComplexMake, args)
   602  		x.commonSelectors[sk] = a
   603  		return a
   604  
   605  	case types.TCOMPLEX128:
   606  		addArg(x.rewriteSelectOrArg(pos, b, container, nil, m0, x.typs.Float64, rc.next(x.typs.Float64)))
   607  		pos = pos.WithNotStmt()
   608  		addArg(x.rewriteSelectOrArg(pos, b, container, nil, m0, x.typs.Float64, rc.next(x.typs.Float64)))
   609  		a = makeOf(a, OpComplexMake, args)
   610  		x.commonSelectors[sk] = a
   611  		return a
   612  
   613  	case types.TINT64:
   614  		if at.Size() > x.regSize {
   615  			addArg(x.rewriteSelectOrArg(pos, b, container, nil, m0, x.firstType, rc.next(x.firstType)))
   616  			pos = pos.WithNotStmt()
   617  			addArg(x.rewriteSelectOrArg(pos, b, container, nil, m0, x.secondType, rc.next(x.secondType)))
   618  			if !x.f.Config.BigEndian {
   619  				// Int64Make args are big, little
   620  				args[0], args[1] = args[1], args[0]
   621  			}
   622  			a = makeOf(a, OpInt64Make, args)
   623  			x.commonSelectors[sk] = a
   624  			return a
   625  		}
   626  	case types.TUINT64:
   627  		if at.Size() > x.regSize {
   628  			addArg(x.rewriteSelectOrArg(pos, b, container, nil, m0, x.typs.UInt32, rc.next(x.typs.UInt32)))
   629  			pos = pos.WithNotStmt()
   630  			addArg(x.rewriteSelectOrArg(pos, b, container, nil, m0, x.typs.UInt32, rc.next(x.typs.UInt32)))
   631  			if !x.f.Config.BigEndian {
   632  				// Int64Make args are big, little
   633  				args[0], args[1] = args[1], args[0]
   634  			}
   635  			a = makeOf(a, OpInt64Make, args)
   636  			x.commonSelectors[sk] = a
   637  			return a
   638  		}
   639  	}
   640  
   641  	// An atomic type, either record the register or store it and update the memory.
   642  
   643  	// Depending on the container Op, the leaves are either OpSelectN or OpArg{Int,Float}Reg
   644  
   645  	if container.Op == OpArg {
   646  		if rc.hasRegs() {
   647  			op, i := rc.ArgOpAndRegisterFor()
   648  			name := container.Aux.(*ir.Name)
   649  			a = makeOf(a, op, nil)
   650  			a.AuxInt = i
   651  			a.Aux = &AuxNameOffset{name, rc.storeOffset}
   652  		} else {
   653  			key := selKey{container, rc.storeOffset, at.Size(), at}
   654  			w := x.commonArgs[key]
   655  			if w != nil && w.Uses != 0 {
   656  				if a == nil {
   657  					a = w
   658  				} else {
   659  					a.copyOf(w)
   660  				}
   661  			} else {
   662  				if a == nil {
   663  					aux := container.Aux
   664  					auxInt := container.AuxInt + rc.storeOffset
   665  					a = container.Block.NewValue0IA(container.Pos, OpArg, at, auxInt, aux)
   666  				} else {
   667  					// do nothing, the original should be okay.
   668  				}
   669  				x.commonArgs[key] = a
   670  			}
   671  		}
   672  	} else if container.Op == OpSelectN {
   673  		call := container.Args[0]
   674  		aux := call.Aux.(*AuxCall)
   675  		which := container.AuxInt
   676  
   677  		if at == types.TypeMem {
   678  			if a != m0 || a != x.memForCall[call.ID] {
   679  				panic(fmt.Errorf("Memories %s, %s, and %s should all be equal after %s", a.LongString(), m0.LongString(), x.memForCall[call.ID], call.LongString()))
   680  			}
   681  		} else if rc.hasRegs() {
   682  			firstReg := uint32(0)
   683  			for i := 0; i < int(which); i++ {
   684  				firstReg += uint32(len(aux.abiInfo.OutParam(i).Registers))
   685  			}
   686  			reg := int64(rc.nextSlice + Abi1RO(firstReg))
   687  			a = makeOf(a, OpSelectN, []*Value{call})
   688  			a.AuxInt = reg
   689  		} else {
   690  			off := x.offsetFrom(x.f.Entry, x.sp, rc.storeOffset+aux.OffsetOfResult(which), types.NewPtr(at))
   691  			a = makeOf(a, OpLoad, []*Value{off, m0})
   692  		}
   693  
   694  	} else {
   695  		panic(fmt.Errorf("Expected container OpArg or OpSelectN, saw %v instead", container.LongString()))
   696  	}
   697  
   698  	x.commonSelectors[sk] = a
   699  	return a
   700  }
   701  
   702  // rewriteWideSelectToStores handles the case of a SelectN'd result from a function call that is too large for SSA,
   703  // but is transferred in registers.  In this case the register cursor tracks both operands; the register sources and
   704  // the memory destinations.
   705  // This returns the memory flowing out of the last store
   706  func (x *expandState) rewriteWideSelectToStores(pos src.XPos, b *Block, container, m0 *Value, at *types.Type, rc registerCursor) *Value {
   707  
   708  	if at.Size() == 0 {
   709  		return m0
   710  	}
   711  
   712  	switch at.Kind() {
   713  	case types.TARRAY:
   714  		et := at.Elem()
   715  		for i := int64(0); i < at.NumElem(); i++ {
   716  			m0 = x.rewriteWideSelectToStores(pos, b, container, m0, et, rc.next(et))
   717  		}
   718  		return m0
   719  
   720  	case types.TSTRUCT:
   721  		// Assume ssagen/ssa.go (in buildssa) spills large aggregates so they won't appear here.
   722  		if at.IsSIMD() {
   723  			break // XXX
   724  		}
   725  		for i := 0; i < at.NumFields(); i++ {
   726  			et := at.Field(i).Type
   727  			m0 = x.rewriteWideSelectToStores(pos, b, container, m0, et, rc.next(et))
   728  			pos = pos.WithNotStmt()
   729  		}
   730  		return m0
   731  
   732  	case types.TSLICE:
   733  		m0 = x.rewriteWideSelectToStores(pos, b, container, m0, at.Elem().PtrTo(), rc.next(x.typs.BytePtr))
   734  		pos = pos.WithNotStmt()
   735  		m0 = x.rewriteWideSelectToStores(pos, b, container, m0, x.typs.Int, rc.next(x.typs.Int))
   736  		m0 = x.rewriteWideSelectToStores(pos, b, container, m0, x.typs.Int, rc.next(x.typs.Int))
   737  		return m0
   738  
   739  	case types.TSTRING:
   740  		m0 = x.rewriteWideSelectToStores(pos, b, container, m0, x.typs.BytePtr, rc.next(x.typs.BytePtr))
   741  		pos = pos.WithNotStmt()
   742  		m0 = x.rewriteWideSelectToStores(pos, b, container, m0, x.typs.Int, rc.next(x.typs.Int))
   743  		return m0
   744  
   745  	case types.TINTER:
   746  		m0 = x.rewriteWideSelectToStores(pos, b, container, m0, x.typs.Uintptr, rc.next(x.typs.Uintptr))
   747  		pos = pos.WithNotStmt()
   748  		m0 = x.rewriteWideSelectToStores(pos, b, container, m0, x.typs.BytePtr, rc.next(x.typs.BytePtr))
   749  		return m0
   750  
   751  	case types.TCOMPLEX64:
   752  		m0 = x.rewriteWideSelectToStores(pos, b, container, m0, x.typs.Float32, rc.next(x.typs.Float32))
   753  		pos = pos.WithNotStmt()
   754  		m0 = x.rewriteWideSelectToStores(pos, b, container, m0, x.typs.Float32, rc.next(x.typs.Float32))
   755  		return m0
   756  
   757  	case types.TCOMPLEX128:
   758  		m0 = x.rewriteWideSelectToStores(pos, b, container, m0, x.typs.Float64, rc.next(x.typs.Float64))
   759  		pos = pos.WithNotStmt()
   760  		m0 = x.rewriteWideSelectToStores(pos, b, container, m0, x.typs.Float64, rc.next(x.typs.Float64))
   761  		return m0
   762  
   763  	case types.TINT64:
   764  		if at.Size() > x.regSize {
   765  			m0 = x.rewriteWideSelectToStores(pos, b, container, m0, x.firstType, rc.next(x.firstType))
   766  			pos = pos.WithNotStmt()
   767  			m0 = x.rewriteWideSelectToStores(pos, b, container, m0, x.secondType, rc.next(x.secondType))
   768  			return m0
   769  		}
   770  	case types.TUINT64:
   771  		if at.Size() > x.regSize {
   772  			m0 = x.rewriteWideSelectToStores(pos, b, container, m0, x.typs.UInt32, rc.next(x.typs.UInt32))
   773  			pos = pos.WithNotStmt()
   774  			m0 = x.rewriteWideSelectToStores(pos, b, container, m0, x.typs.UInt32, rc.next(x.typs.UInt32))
   775  			return m0
   776  		}
   777  	}
   778  
   779  	// TODO could change treatment of too-large OpArg, would deal with it here.
   780  	if container.Op == OpSelectN {
   781  		call := container.Args[0]
   782  		aux := call.Aux.(*AuxCall)
   783  		which := container.AuxInt
   784  
   785  		if rc.hasRegs() {
   786  			firstReg := uint32(0)
   787  			for i := 0; i < int(which); i++ {
   788  				firstReg += uint32(len(aux.abiInfo.OutParam(i).Registers))
   789  			}
   790  			reg := int64(rc.nextSlice + Abi1RO(firstReg))
   791  			a := b.NewValue1I(pos, OpSelectN, at, reg, call)
   792  			dst := x.offsetFrom(b, rc.storeDest, rc.storeOffset, types.NewPtr(at))
   793  			m0 = b.NewValue3A(pos, OpStore, types.TypeMem, at, dst, a, m0)
   794  		} else {
   795  			panic(fmt.Errorf("Expected rc to have registers"))
   796  		}
   797  	} else {
   798  		panic(fmt.Errorf("Expected container OpSelectN, saw %v instead", container.LongString()))
   799  	}
   800  	return m0
   801  }
   802  
   803  func isBlockMultiValueExit(b *Block) bool {
   804  	return (b.Kind == BlockRet || b.Kind == BlockRetJmp) && b.Controls[0] != nil && b.Controls[0].Op == OpMakeResult
   805  }
   806  
   807  type Abi1RO uint8 // An offset within a parameter's slice of register indices, for abi1.
   808  
   809  // A registerCursor tracks which register is used for an Arg or regValues, or a piece of such.
   810  type registerCursor struct {
   811  	storeDest   *Value // if there are no register targets, then this is the base of the store.
   812  	storeOffset int64
   813  	regs        []abi.RegIndex // the registers available for this Arg/result (which is all in registers or not at all)
   814  	nextSlice   Abi1RO         // the next register/register-slice offset
   815  	config      *abi.ABIConfig
   816  	regValues   *[]*Value // values assigned to registers accumulate here
   817  }
   818  
   819  func (c *registerCursor) String() string {
   820  	dest := "<none>"
   821  	if c.storeDest != nil {
   822  		dest = fmt.Sprintf("%s+%d", c.storeDest.String(), c.storeOffset)
   823  	}
   824  	regs := "<none>"
   825  	if c.regValues != nil {
   826  		regs = ""
   827  		for i, x := range *c.regValues {
   828  			if i > 0 {
   829  				regs = regs + "; "
   830  			}
   831  			regs = regs + x.LongString()
   832  		}
   833  	}
   834  
   835  	// not printing the config because that has not been useful
   836  	return fmt.Sprintf("RCSR{storeDest=%v, regsLen=%d, nextSlice=%d, regValues=[%s]}", dest, len(c.regs), c.nextSlice, regs)
   837  }
   838  
   839  // next effectively post-increments the register cursor; the receiver is advanced,
   840  // the (aligned) old value is returned.
   841  func (c *registerCursor) next(t *types.Type) registerCursor {
   842  	c.storeOffset = types.RoundUp(c.storeOffset, t.Alignment())
   843  	rc := *c
   844  	c.storeOffset = types.RoundUp(c.storeOffset+t.Size(), t.Alignment())
   845  	if int(c.nextSlice) < len(c.regs) {
   846  		w := c.config.NumParamRegs(t)
   847  		c.nextSlice += Abi1RO(w)
   848  	}
   849  	return rc
   850  }
   851  
   852  // plus returns a register cursor offset from the original, without modifying the original.
   853  func (c *registerCursor) plus(regWidth Abi1RO) registerCursor {
   854  	rc := *c
   855  	rc.nextSlice += regWidth
   856  	return rc
   857  }
   858  
   859  func (c *registerCursor) init(regs []abi.RegIndex, info *abi.ABIParamResultInfo, result *[]*Value, storeDest *Value, storeOffset int64) {
   860  	c.regs = regs
   861  	c.nextSlice = 0
   862  	c.storeOffset = storeOffset
   863  	c.storeDest = storeDest
   864  	c.config = info.Config()
   865  	c.regValues = result
   866  }
   867  
   868  func (c *registerCursor) addArg(v *Value) {
   869  	*c.regValues = append(*c.regValues, v)
   870  }
   871  
   872  func (c *registerCursor) hasRegs() bool {
   873  	return len(c.regs) > 0
   874  }
   875  
   876  func (c *registerCursor) ArgOpAndRegisterFor() (Op, int64) {
   877  	r := c.regs[c.nextSlice]
   878  	return ArgOpAndRegisterFor(r, c.config)
   879  }
   880  
   881  // ArgOpAndRegisterFor converts an abi register index into an ssa Op and corresponding
   882  // arg register index.
   883  func ArgOpAndRegisterFor(r abi.RegIndex, abiConfig *abi.ABIConfig) (Op, int64) {
   884  	i := abiConfig.FloatIndexFor(r)
   885  	if i >= 0 { // float PR
   886  		return OpArgFloatReg, i
   887  	}
   888  	return OpArgIntReg, int64(r)
   889  }
   890  
   891  type selKey struct {
   892  	from          *Value // what is selected from
   893  	offsetOrIndex int64  // whatever is appropriate for the selector
   894  	size          int64
   895  	typ           *types.Type
   896  }
   897  
   898  type expandState struct {
   899  	f       *Func
   900  	debug   int // odd values log lost statement markers, so likely settings are 1 (stmts), 2 (expansion), and 3 (both)
   901  	regSize int64
   902  	sp      *Value
   903  	typs    *Types
   904  
   905  	firstOp    Op          // for 64-bit integers on 32-bit machines, first word in memory
   906  	secondOp   Op          // for 64-bit integers on 32-bit machines, second word in memory
   907  	firstType  *types.Type // first half type, for Int64
   908  	secondType *types.Type // second half type, for Int64
   909  
   910  	wideSelects     map[*Value]*Value // Selects that are not SSA-able, mapped to consuming stores.
   911  	commonSelectors map[selKey]*Value // used to de-dupe selectors
   912  	commonArgs      map[selKey]*Value // used to de-dupe OpArg/OpArgIntReg/OpArgFloatReg
   913  	memForCall      map[ID]*Value     // For a call, need to know the unique selector that gets the mem.
   914  	indentLevel     int               // Indentation for debugging recursion
   915  }
   916  
   917  // offsetFrom creates an offset from a pointer, simplifying chained offsets and offsets from SP
   918  func (x *expandState) offsetFrom(b *Block, from *Value, offset int64, pt *types.Type) *Value {
   919  	ft := from.Type
   920  	if offset == 0 {
   921  		if ft == pt {
   922  			return from
   923  		}
   924  		// This captures common, (apparently) safe cases.  The unsafe cases involve ft == uintptr
   925  		if (ft.IsPtr() || ft.IsUnsafePtr()) && pt.IsPtr() {
   926  			return from
   927  		}
   928  	}
   929  	// Simplify, canonicalize
   930  	for from.Op == OpOffPtr {
   931  		offset += from.AuxInt
   932  		from = from.Args[0]
   933  	}
   934  	if from == x.sp {
   935  		return x.f.ConstOffPtrSP(pt, offset, x.sp)
   936  	}
   937  	return b.NewValue1I(from.Pos.WithNotStmt(), OpOffPtr, pt, offset, from)
   938  }
   939  
   940  // prAssignForArg returns the ABIParamAssignment for v, assumed to be an OpArg.
   941  func (x *expandState) prAssignForArg(v *Value) *abi.ABIParamAssignment {
   942  	if v.Op != OpArg {
   943  		panic(fmt.Errorf("Wanted OpArg, instead saw %s", v.LongString()))
   944  	}
   945  	return ParamAssignmentForArgName(x.f, v.Aux.(*ir.Name))
   946  }
   947  
   948  // ParamAssignmentForArgName returns the ABIParamAssignment for f's arg with matching name.
   949  func ParamAssignmentForArgName(f *Func, name *ir.Name) *abi.ABIParamAssignment {
   950  	abiInfo := f.OwnAux.abiInfo
   951  	ip := abiInfo.InParams()
   952  	for i, a := range ip {
   953  		if a.Name == name {
   954  			return &ip[i]
   955  		}
   956  	}
   957  	panic(fmt.Errorf("Did not match param %v in prInfo %+v", name, abiInfo.InParams()))
   958  }
   959  
   960  // indent increments (or decrements) the indentation.
   961  func (x *expandState) indent(n int) {
   962  	x.indentLevel += n
   963  }
   964  
   965  // Printf does an indented fmt.Printf on the format and args.
   966  func (x *expandState) Printf(format string, a ...any) (n int, err error) {
   967  	if x.indentLevel > 0 {
   968  		fmt.Printf("%[1]*s", x.indentLevel, "")
   969  	}
   970  	return fmt.Printf(format, a...)
   971  }
   972  
   973  func (x *expandState) invalidateRecursively(a *Value) {
   974  	var s string
   975  	if x.debug > 0 {
   976  		plus := " "
   977  		if a.Pos.IsStmt() == src.PosIsStmt {
   978  			plus = " +"
   979  		}
   980  		s = a.String() + plus + a.Pos.LineNumber() + " " + a.LongString()
   981  		if x.debug > 1 {
   982  			x.Printf("...marking %v unused\n", s)
   983  		}
   984  	}
   985  	lost := a.invalidateRecursively()
   986  	if x.debug&1 != 0 && lost { // For odd values of x.debug, do this.
   987  		x.Printf("Lost statement marker in %s on former %s\n", base.Ctxt.Pkgpath+"."+x.f.Name, s)
   988  	}
   989  }
   990  

View as plain text