Source file src/cmd/compile/internal/ssa/_gen/rulegen.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  // This program generates Go code that applies rewrite rules to a Value.
     6  // The generated code implements a function of type func (v *Value) bool
     7  // which reports whether if did something.
     8  // Ideas stolen from the Swift Java compiler:
     9  // https://bitsavers.org/pdf/dec/tech_reports/WRL-2000-2.pdf
    10  
    11  package main
    12  
    13  import (
    14  	"bufio"
    15  	"bytes"
    16  	"flag"
    17  	"fmt"
    18  	"go/ast"
    19  	"go/format"
    20  	"go/parser"
    21  	"go/printer"
    22  	"go/token"
    23  	"io"
    24  	"log"
    25  	"os"
    26  	"path"
    27  	"regexp"
    28  	"sort"
    29  	"strconv"
    30  	"strings"
    31  
    32  	"golang.org/x/tools/go/ast/astutil"
    33  )
    34  
    35  // rule syntax:
    36  //  sexpr [&& extra conditions] => [@block] sexpr
    37  //
    38  // sexpr are s-expressions (lisp-like parenthesized groupings)
    39  // sexpr ::= [variable:](opcode sexpr*)
    40  //         | variable
    41  //         | <type>
    42  //         | [auxint]
    43  //         | {aux}
    44  //
    45  // aux      ::= variable | {code}
    46  // type     ::= variable | {code}
    47  // variable ::= some token
    48  // opcode   ::= one of the opcodes from the *Ops.go files
    49  
    50  // special rules: trailing ellipsis "..." (in the outermost sexpr?) must match on both sides of a rule.
    51  //                trailing three underscore "___" in the outermost match sexpr indicate the presence of
    52  //                   extra ignored args that need not appear in the replacement
    53  //                if the right-hand side is in {}, then it is code used to generate the result.
    54  
    55  // extra conditions is just a chunk of Go that evaluates to a boolean. It may use
    56  // variables declared in the matching tsexpr. The variable "v" is predefined to be
    57  // the value matched by the entire rule.
    58  
    59  // If multiple rules match, the first one in file order is selected.
    60  
    61  var (
    62  	genLog  = flag.Bool("log", false, "generate code that logs; for debugging only")
    63  	addLine = flag.Bool("line", false, "add line number comment to generated rules; for debugging only")
    64  )
    65  
    66  type Rule struct {
    67  	Rule string
    68  	Loc  string // file name & line number
    69  }
    70  
    71  func (r Rule) String() string {
    72  	return fmt.Sprintf("rule %q at %s", r.Rule, r.Loc)
    73  }
    74  
    75  func normalizeSpaces(s string) string {
    76  	return strings.Join(strings.Fields(strings.TrimSpace(s)), " ")
    77  }
    78  
    79  // parse returns the matching part of the rule, additional conditions, and the result.
    80  func (r Rule) parse() (match, cond, result string) {
    81  	s := strings.Split(r.Rule, "=>")
    82  	match = normalizeSpaces(s[0])
    83  	result = normalizeSpaces(s[1])
    84  	cond = ""
    85  	if i := strings.Index(match, "&&"); i >= 0 {
    86  		cond = normalizeSpaces(match[i+2:])
    87  		match = normalizeSpaces(match[:i])
    88  	}
    89  	return match, cond, result
    90  }
    91  
    92  func genRules(arch arch)          { genRulesSuffix(arch, "") }
    93  func genSplitLoadRules(arch arch) { genRulesSuffix(arch, "splitload") }
    94  func genLateLowerRules(arch arch) { genRulesSuffix(arch, "latelower") }
    95  
    96  func genRulesSuffix(arch arch, suff string) {
    97  	// Open input file.
    98  	text, err := os.Open(arch.name + suff + ".rules")
    99  	if err != nil {
   100  		if suff == "" {
   101  			// All architectures must have a plain rules file.
   102  			log.Fatalf("can't read rule file: %v", err)
   103  		}
   104  		// Some architectures have bonus rules files that others don't share. That's fine.
   105  		return
   106  	}
   107  
   108  	// oprules contains a list of rules for each block and opcode
   109  	blockrules := map[string][]Rule{}
   110  	oprules := map[string][]Rule{}
   111  
   112  	// read rule file
   113  	scanner := bufio.NewScanner(text)
   114  	rule := ""
   115  	var lineno int
   116  	var ruleLineno int // line number of "=>"
   117  	for scanner.Scan() {
   118  		lineno++
   119  		line := scanner.Text()
   120  		if i := strings.Index(line, "//"); i >= 0 {
   121  			// Remove comments. Note that this isn't string safe, so
   122  			// it will truncate lines with // inside strings. Oh well.
   123  			line = line[:i]
   124  		}
   125  		rule += " " + line
   126  		rule = strings.TrimSpace(rule)
   127  		if rule == "" {
   128  			continue
   129  		}
   130  		if !strings.Contains(rule, "=>") {
   131  			continue
   132  		}
   133  		if ruleLineno == 0 {
   134  			ruleLineno = lineno
   135  		}
   136  		if strings.HasSuffix(rule, "=>") {
   137  			continue // continue on the next line
   138  		}
   139  		if n := balance(rule); n > 0 {
   140  			continue // open parentheses remain, continue on the next line
   141  		} else if n < 0 {
   142  			break // continuing the line can't help, and it will only make errors worse
   143  		}
   144  
   145  		loc := fmt.Sprintf("%s%s.rules:%d", arch.name, suff, ruleLineno)
   146  		for _, rule2 := range expandOr(rule) {
   147  			r := Rule{Rule: rule2, Loc: loc}
   148  			if rawop := strings.Split(rule2, " ")[0][1:]; isBlock(rawop, arch) {
   149  				blockrules[rawop] = append(blockrules[rawop], r)
   150  				continue
   151  			}
   152  			// Do fancier value op matching.
   153  			match, _, _ := r.parse()
   154  			op, oparch, _, _, _, _ := parseValue(match, arch, loc)
   155  			opname := fmt.Sprintf("Op%s%s", oparch, op.name)
   156  			oprules[opname] = append(oprules[opname], r)
   157  		}
   158  		rule = ""
   159  		ruleLineno = 0
   160  	}
   161  	if err := scanner.Err(); err != nil {
   162  		log.Fatalf("scanner failed: %v\n", err)
   163  	}
   164  	if balance(rule) != 0 {
   165  		log.Fatalf("%s.rules:%d: unbalanced rule: %v\n", arch.name, lineno, rule)
   166  	}
   167  
   168  	// Order all the ops.
   169  	var ops []string
   170  	for op := range oprules {
   171  		ops = append(ops, op)
   172  	}
   173  	sort.Strings(ops)
   174  
   175  	genFile := &File{Arch: arch, Suffix: suff}
   176  	// Main rewrite routine is a switch on v.Op.
   177  	fn := &Func{Kind: "Value", ArgLen: -1}
   178  
   179  	sw := &Switch{Expr: exprf("v.Op")}
   180  	for _, op := range ops {
   181  		eop, ok := parseEllipsisRules(oprules[op], arch)
   182  		if ok {
   183  			if strings.Contains(oprules[op][0].Rule, "=>") && opByName(arch, op).aux != opByName(arch, eop).aux {
   184  				panic(fmt.Sprintf("can't use ... for ops that have different aux types: %s and %s", op, eop))
   185  			}
   186  			swc := &Case{Expr: exprf("%s", op)}
   187  			swc.add(stmtf("v.Op = %s", eop))
   188  			swc.add(stmtf("return true"))
   189  			sw.add(swc)
   190  			continue
   191  		}
   192  
   193  		swc := &Case{Expr: exprf("%s", op)}
   194  		swc.add(stmtf("return rewriteValue%s%s_%s(v)", arch.name, suff, op))
   195  		sw.add(swc)
   196  	}
   197  	if len(sw.List) > 0 { // skip if empty
   198  		fn.add(sw)
   199  	}
   200  	fn.add(stmtf("return false"))
   201  	genFile.add(fn)
   202  
   203  	// Generate a routine per op. Note that we don't make one giant routine
   204  	// because it is too big for some compilers.
   205  	for _, op := range ops {
   206  		rules := oprules[op]
   207  		_, ok := parseEllipsisRules(oprules[op], arch)
   208  		if ok {
   209  			continue
   210  		}
   211  
   212  		// rr is kept between iterations, so that each rule can check
   213  		// that the previous rule wasn't unconditional.
   214  		var rr *RuleRewrite
   215  		fn := &Func{
   216  			Kind:   "Value",
   217  			Suffix: fmt.Sprintf("_%s", op),
   218  			ArgLen: opByName(arch, op).argLength,
   219  		}
   220  		fn.add(declReserved("b", "v.Block"))
   221  		fn.add(declReserved("config", "b.Func.Config"))
   222  		fn.add(declReserved("fe", "b.Func.fe"))
   223  		fn.add(declReserved("typ", "&b.Func.Config.Types"))
   224  		for _, rule := range rules {
   225  			if rr != nil && !rr.CanFail {
   226  				log.Fatalf("unconditional rule %s is followed by other rules", rr.Match)
   227  			}
   228  			rr = &RuleRewrite{Loc: rule.Loc}
   229  			rr.Match, rr.Cond, rr.Result = rule.parse()
   230  			pos, _ := genMatch(rr, arch, rr.Match, fn.ArgLen >= 0)
   231  			if pos == "" {
   232  				pos = "v.Pos"
   233  			}
   234  			if rr.Cond != "" {
   235  				rr.add(breakf("!(%s)", rr.Cond))
   236  			}
   237  			genResult(rr, arch, rr.Result, pos)
   238  			if *genLog {
   239  				rr.add(stmtf("logRule(%q)", rule.Loc))
   240  			}
   241  			fn.add(rr)
   242  		}
   243  		if rr.CanFail {
   244  			fn.add(stmtf("return false"))
   245  		}
   246  		genFile.add(fn)
   247  	}
   248  
   249  	// Generate block rewrite function. There are only a few block types
   250  	// so we can make this one function with a switch.
   251  	fn = &Func{Kind: "Block"}
   252  	fn.add(declReserved("config", "b.Func.Config"))
   253  	fn.add(declReserved("typ", "&b.Func.Config.Types"))
   254  
   255  	sw = &Switch{Expr: exprf("b.Kind")}
   256  	ops = ops[:0]
   257  	for op := range blockrules {
   258  		ops = append(ops, op)
   259  	}
   260  	sort.Strings(ops)
   261  	for _, op := range ops {
   262  		name, data := getBlockInfo(op, arch)
   263  		swc := &Case{Expr: exprf("%s", name)}
   264  		for _, rule := range blockrules[op] {
   265  			swc.add(genBlockRewrite(rule, arch, data))
   266  		}
   267  		sw.add(swc)
   268  	}
   269  	if len(sw.List) > 0 { // skip if empty
   270  		fn.add(sw)
   271  	}
   272  	fn.add(stmtf("return false"))
   273  	genFile.add(fn)
   274  
   275  	// Remove unused imports and variables.
   276  	buf := new(bytes.Buffer)
   277  	fprint(buf, genFile)
   278  	fset := token.NewFileSet()
   279  	file, err := parser.ParseFile(fset, "", buf, parser.ParseComments)
   280  	if err != nil {
   281  		filename := fmt.Sprintf("%s_broken.go", arch.name)
   282  		if err := os.WriteFile(filename, buf.Bytes(), 0644); err != nil {
   283  			log.Printf("failed to dump broken code to %s: %v", filename, err)
   284  		} else {
   285  			log.Printf("dumped broken code to %s", filename)
   286  		}
   287  		log.Fatalf("failed to parse generated code for arch %s: %v", arch.name, err)
   288  	}
   289  	tfile := fset.File(file.Pos())
   290  
   291  	// First, use unusedInspector to find the unused declarations by their
   292  	// start position.
   293  	u := unusedInspector{unused: make(map[token.Pos]bool)}
   294  	u.node(file)
   295  
   296  	// Then, delete said nodes via astutil.Apply.
   297  	pre := func(c *astutil.Cursor) bool {
   298  		node := c.Node()
   299  		if node == nil {
   300  			return true
   301  		}
   302  		if u.unused[node.Pos()] {
   303  			c.Delete()
   304  			// Unused imports and declarations use exactly
   305  			// one line. Prevent leaving an empty line.
   306  			tfile.MergeLine(tfile.Position(node.Pos()).Line)
   307  			return false
   308  		}
   309  		return true
   310  	}
   311  	post := func(c *astutil.Cursor) bool {
   312  		switch node := c.Node().(type) {
   313  		case *ast.GenDecl:
   314  			if len(node.Specs) == 0 {
   315  				// Don't leave a broken or empty GenDecl behind,
   316  				// such as "import ()".
   317  				c.Delete()
   318  			}
   319  		}
   320  		return true
   321  	}
   322  	file = astutil.Apply(file, pre, post).(*ast.File)
   323  
   324  	// Write the well-formatted source to file
   325  	f, err := os.Create("../rewrite" + arch.name + suff + ".go")
   326  	if err != nil {
   327  		log.Fatalf("can't write output: %v", err)
   328  	}
   329  	defer f.Close()
   330  	// gofmt result; use a buffered writer, as otherwise go/format spends
   331  	// far too much time in syscalls.
   332  	bw := bufio.NewWriter(f)
   333  	if err := format.Node(bw, fset, file); err != nil {
   334  		log.Fatalf("can't format output: %v", err)
   335  	}
   336  	if err := bw.Flush(); err != nil {
   337  		log.Fatalf("can't write output: %v", err)
   338  	}
   339  	if err := f.Close(); err != nil {
   340  		log.Fatalf("can't write output: %v", err)
   341  	}
   342  }
   343  
   344  // unusedInspector can be used to detect unused variables and imports in an
   345  // ast.Node via its node method. The result is available in the "unused" map.
   346  //
   347  // note that unusedInspector is lazy and best-effort; it only supports the node
   348  // types and patterns used by the rulegen program.
   349  type unusedInspector struct {
   350  	// scope is the current scope, which can never be nil when a declaration
   351  	// is encountered. That is, the unusedInspector.node entrypoint should
   352  	// generally be an entire file or block.
   353  	scope *scope
   354  
   355  	// unused is the resulting set of unused declared names, indexed by the
   356  	// starting position of the node that declared the name.
   357  	unused map[token.Pos]bool
   358  
   359  	// defining is the object currently being defined; this is useful so
   360  	// that if "foo := bar" is unused and removed, we can then detect if
   361  	// "bar" becomes unused as well.
   362  	defining *object
   363  }
   364  
   365  // scoped opens a new scope when called, and returns a function which closes
   366  // that same scope. When a scope is closed, unused variables are recorded.
   367  func (u *unusedInspector) scoped() func() {
   368  	outer := u.scope
   369  	u.scope = &scope{outer: outer, objects: map[string]*object{}}
   370  	return func() {
   371  		for anyUnused := true; anyUnused; {
   372  			anyUnused = false
   373  			for _, obj := range u.scope.objects {
   374  				if obj.numUses > 0 {
   375  					continue
   376  				}
   377  				u.unused[obj.pos] = true
   378  				for _, used := range obj.used {
   379  					if used.numUses--; used.numUses == 0 {
   380  						anyUnused = true
   381  					}
   382  				}
   383  				// We've decremented numUses for each of the
   384  				// objects in used. Zero this slice too, to keep
   385  				// everything consistent.
   386  				obj.used = nil
   387  			}
   388  		}
   389  		u.scope = outer
   390  	}
   391  }
   392  
   393  func (u *unusedInspector) exprs(list []ast.Expr) {
   394  	for _, x := range list {
   395  		u.node(x)
   396  	}
   397  }
   398  
   399  func (u *unusedInspector) node(node ast.Node) {
   400  	switch node := node.(type) {
   401  	case *ast.File:
   402  		defer u.scoped()()
   403  		for _, decl := range node.Decls {
   404  			u.node(decl)
   405  		}
   406  	case *ast.GenDecl:
   407  		for _, spec := range node.Specs {
   408  			u.node(spec)
   409  		}
   410  	case *ast.ImportSpec:
   411  		impPath, _ := strconv.Unquote(node.Path.Value)
   412  		name := path.Base(impPath)
   413  		u.scope.objects[name] = &object{
   414  			name: name,
   415  			pos:  node.Pos(),
   416  		}
   417  	case *ast.FuncDecl:
   418  		u.node(node.Type)
   419  		if node.Body != nil {
   420  			u.node(node.Body)
   421  		}
   422  	case *ast.FuncType:
   423  		if node.Params != nil {
   424  			u.node(node.Params)
   425  		}
   426  		if node.Results != nil {
   427  			u.node(node.Results)
   428  		}
   429  	case *ast.FieldList:
   430  		for _, field := range node.List {
   431  			u.node(field)
   432  		}
   433  	case *ast.Field:
   434  		u.node(node.Type)
   435  
   436  	// statements
   437  
   438  	case *ast.BlockStmt:
   439  		defer u.scoped()()
   440  		for _, stmt := range node.List {
   441  			u.node(stmt)
   442  		}
   443  	case *ast.DeclStmt:
   444  		u.node(node.Decl)
   445  	case *ast.IfStmt:
   446  		if node.Init != nil {
   447  			u.node(node.Init)
   448  		}
   449  		u.node(node.Cond)
   450  		u.node(node.Body)
   451  		if node.Else != nil {
   452  			u.node(node.Else)
   453  		}
   454  	case *ast.ForStmt:
   455  		if node.Init != nil {
   456  			u.node(node.Init)
   457  		}
   458  		if node.Cond != nil {
   459  			u.node(node.Cond)
   460  		}
   461  		if node.Post != nil {
   462  			u.node(node.Post)
   463  		}
   464  		u.node(node.Body)
   465  	case *ast.SwitchStmt:
   466  		if node.Init != nil {
   467  			u.node(node.Init)
   468  		}
   469  		if node.Tag != nil {
   470  			u.node(node.Tag)
   471  		}
   472  		u.node(node.Body)
   473  	case *ast.CaseClause:
   474  		u.exprs(node.List)
   475  		defer u.scoped()()
   476  		for _, stmt := range node.Body {
   477  			u.node(stmt)
   478  		}
   479  	case *ast.BranchStmt:
   480  	case *ast.ExprStmt:
   481  		u.node(node.X)
   482  	case *ast.AssignStmt:
   483  		if node.Tok != token.DEFINE {
   484  			u.exprs(node.Rhs)
   485  			u.exprs(node.Lhs)
   486  			break
   487  		}
   488  		lhs := node.Lhs
   489  		if len(lhs) == 2 && lhs[1].(*ast.Ident).Name == "_" {
   490  			lhs = lhs[:1]
   491  		}
   492  		if len(lhs) != 1 {
   493  			panic("no support for := with multiple names")
   494  		}
   495  
   496  		name := lhs[0].(*ast.Ident)
   497  		obj := &object{
   498  			name: name.Name,
   499  			pos:  name.NamePos,
   500  		}
   501  
   502  		old := u.defining
   503  		u.defining = obj
   504  		u.exprs(node.Rhs)
   505  		u.defining = old
   506  
   507  		u.scope.objects[name.Name] = obj
   508  	case *ast.ReturnStmt:
   509  		u.exprs(node.Results)
   510  	case *ast.IncDecStmt:
   511  		u.node(node.X)
   512  
   513  	// expressions
   514  
   515  	case *ast.CallExpr:
   516  		u.node(node.Fun)
   517  		u.exprs(node.Args)
   518  	case *ast.SelectorExpr:
   519  		u.node(node.X)
   520  	case *ast.UnaryExpr:
   521  		u.node(node.X)
   522  	case *ast.BinaryExpr:
   523  		u.node(node.X)
   524  		u.node(node.Y)
   525  	case *ast.StarExpr:
   526  		u.node(node.X)
   527  	case *ast.ParenExpr:
   528  		u.node(node.X)
   529  	case *ast.IndexExpr:
   530  		u.node(node.X)
   531  		u.node(node.Index)
   532  	case *ast.TypeAssertExpr:
   533  		u.node(node.X)
   534  		u.node(node.Type)
   535  	case *ast.Ident:
   536  		if obj := u.scope.Lookup(node.Name); obj != nil {
   537  			obj.numUses++
   538  			if u.defining != nil {
   539  				u.defining.used = append(u.defining.used, obj)
   540  			}
   541  		}
   542  	case *ast.BasicLit:
   543  	case *ast.ValueSpec:
   544  		u.exprs(node.Values)
   545  	default:
   546  		panic(fmt.Sprintf("unhandled node: %T", node))
   547  	}
   548  }
   549  
   550  // scope keeps track of a certain scope and its declared names, as well as the
   551  // outer (parent) scope.
   552  type scope struct {
   553  	outer   *scope             // can be nil, if this is the top-level scope
   554  	objects map[string]*object // indexed by each declared name
   555  }
   556  
   557  func (s *scope) Lookup(name string) *object {
   558  	if obj := s.objects[name]; obj != nil {
   559  		return obj
   560  	}
   561  	if s.outer == nil {
   562  		return nil
   563  	}
   564  	return s.outer.Lookup(name)
   565  }
   566  
   567  // object keeps track of a declared name, such as a variable or import.
   568  type object struct {
   569  	name string
   570  	pos  token.Pos // start position of the node declaring the object
   571  
   572  	numUses int       // number of times this object is used
   573  	used    []*object // objects that its declaration makes use of
   574  }
   575  
   576  func fprint(w io.Writer, n Node) {
   577  	switch n := n.(type) {
   578  	case *File:
   579  		file := n
   580  		seenRewrite := make(map[[3]string]string)
   581  		fmt.Fprintf(w, "// Code generated from _gen/%s%s.rules using 'go generate'; DO NOT EDIT.\n", n.Arch.name, n.Suffix)
   582  		fmt.Fprintf(w, "\npackage ssa\n")
   583  		for _, path := range append([]string{
   584  			"fmt",
   585  			"internal/buildcfg",
   586  			"math",
   587  			"math/bits",
   588  			"cmd/internal/obj",
   589  			"cmd/compile/internal/base",
   590  			"cmd/compile/internal/types",
   591  			"cmd/compile/internal/ir",
   592  		}, n.Arch.imports...) {
   593  			fmt.Fprintf(w, "import %q\n", path)
   594  		}
   595  		for _, f := range n.List {
   596  			f := f.(*Func)
   597  			fmt.Fprintf(w, "func rewrite%s%s%s%s(", f.Kind, n.Arch.name, n.Suffix, f.Suffix)
   598  			fmt.Fprintf(w, "%c *%s) bool {\n", strings.ToLower(f.Kind)[0], f.Kind)
   599  			if f.Kind == "Value" && f.ArgLen > 0 {
   600  				for i := f.ArgLen - 1; i >= 0; i-- {
   601  					fmt.Fprintf(w, "v_%d := v.Args[%d]\n", i, i)
   602  				}
   603  			}
   604  			for _, n := range f.List {
   605  				fprint(w, n)
   606  
   607  				if rr, ok := n.(*RuleRewrite); ok {
   608  					k := [3]string{
   609  						normalizeMatch(rr.Match, file.Arch),
   610  						normalizeWhitespace(rr.Cond),
   611  						normalizeWhitespace(rr.Result),
   612  					}
   613  					if prev, ok := seenRewrite[k]; ok {
   614  						log.Fatalf("duplicate rule %s, previously seen at %s\n", rr.Loc, prev)
   615  					}
   616  					seenRewrite[k] = rr.Loc
   617  				}
   618  			}
   619  			fmt.Fprintf(w, "}\n")
   620  		}
   621  	case *Switch:
   622  		fmt.Fprintf(w, "switch ")
   623  		fprint(w, n.Expr)
   624  		fmt.Fprintf(w, " {\n")
   625  		for _, n := range n.List {
   626  			fprint(w, n)
   627  		}
   628  		fmt.Fprintf(w, "}\n")
   629  	case *Case:
   630  		fmt.Fprintf(w, "case ")
   631  		fprint(w, n.Expr)
   632  		fmt.Fprintf(w, ":\n")
   633  		for _, n := range n.List {
   634  			fprint(w, n)
   635  		}
   636  	case *RuleRewrite:
   637  		if *addLine {
   638  			fmt.Fprintf(w, "// %s\n", n.Loc)
   639  		}
   640  		fmt.Fprintf(w, "// match: %s\n", n.Match)
   641  		if n.Cond != "" {
   642  			fmt.Fprintf(w, "// cond: %s\n", n.Cond)
   643  		}
   644  		fmt.Fprintf(w, "// result: %s\n", n.Result)
   645  		fmt.Fprintf(w, "for %s {\n", n.Check)
   646  		nCommutative := 0
   647  		for _, n := range n.List {
   648  			if b, ok := n.(*CondBreak); ok {
   649  				b.InsideCommuteLoop = nCommutative > 0
   650  			}
   651  			fprint(w, n)
   652  			if loop, ok := n.(StartCommuteLoop); ok {
   653  				if nCommutative != loop.Depth {
   654  					panic("mismatch commute loop depth")
   655  				}
   656  				nCommutative++
   657  			}
   658  		}
   659  		fmt.Fprintf(w, "return true\n")
   660  		for i := 0; i < nCommutative; i++ {
   661  			fmt.Fprintln(w, "}")
   662  		}
   663  		if n.CommuteDepth > 0 && n.CanFail {
   664  			fmt.Fprint(w, "break\n")
   665  		}
   666  		fmt.Fprintf(w, "}\n")
   667  	case *Declare:
   668  		fmt.Fprintf(w, "%s := ", n.Name)
   669  		fprint(w, n.Value)
   670  		fmt.Fprintln(w)
   671  	case *CondBreak:
   672  		fmt.Fprintf(w, "if ")
   673  		fprint(w, n.Cond)
   674  		fmt.Fprintf(w, " {\n")
   675  		if n.InsideCommuteLoop {
   676  			fmt.Fprintf(w, "continue")
   677  		} else {
   678  			fmt.Fprintf(w, "break")
   679  		}
   680  		fmt.Fprintf(w, "\n}\n")
   681  	case ast.Node:
   682  		printConfig.Fprint(w, emptyFset, n)
   683  		if _, ok := n.(ast.Stmt); ok {
   684  			fmt.Fprintln(w)
   685  		}
   686  	case StartCommuteLoop:
   687  		fmt.Fprintf(w, "for _i%[1]d := 0; _i%[1]d <= 1; _i%[1]d, %[2]s_0, %[2]s_1 = _i%[1]d + 1, %[2]s_1, %[2]s_0 {\n", n.Depth, n.V)
   688  	default:
   689  		log.Fatalf("cannot print %T", n)
   690  	}
   691  }
   692  
   693  var printConfig = printer.Config{
   694  	Mode: printer.RawFormat, // we use go/format later, so skip work here
   695  }
   696  
   697  var emptyFset = token.NewFileSet()
   698  
   699  // Node can be a Statement or an ast.Expr.
   700  type Node interface{}
   701  
   702  // Statement can be one of our high-level statement struct types, or an
   703  // ast.Stmt under some limited circumstances.
   704  type Statement interface{}
   705  
   706  // BodyBase is shared by all of our statement pseudo-node types which can
   707  // contain other statements.
   708  type BodyBase struct {
   709  	List    []Statement
   710  	CanFail bool
   711  }
   712  
   713  func (w *BodyBase) add(node Statement) {
   714  	var last Statement
   715  	if len(w.List) > 0 {
   716  		last = w.List[len(w.List)-1]
   717  	}
   718  	if node, ok := node.(*CondBreak); ok {
   719  		w.CanFail = true
   720  		if last, ok := last.(*CondBreak); ok {
   721  			// Add to the previous "if <cond> { break }" via a
   722  			// logical OR, which will save verbosity.
   723  			last.Cond = &ast.BinaryExpr{
   724  				Op: token.LOR,
   725  				X:  last.Cond,
   726  				Y:  node.Cond,
   727  			}
   728  			return
   729  		}
   730  	}
   731  
   732  	w.List = append(w.List, node)
   733  }
   734  
   735  // predeclared contains globally known tokens that should not be redefined.
   736  var predeclared = map[string]bool{
   737  	"nil":   true,
   738  	"false": true,
   739  	"true":  true,
   740  }
   741  
   742  // declared reports if the body contains a Declare with the given name.
   743  func (w *BodyBase) declared(name string) bool {
   744  	if predeclared[name] {
   745  		// Treat predeclared names as having already been declared.
   746  		// This lets us use nil to match an aux field or
   747  		// true and false to match an auxint field.
   748  		return true
   749  	}
   750  	for _, s := range w.List {
   751  		if decl, ok := s.(*Declare); ok && decl.Name == name {
   752  			return true
   753  		}
   754  	}
   755  	return false
   756  }
   757  
   758  // These types define some high-level statement struct types, which can be used
   759  // as a Statement. This allows us to keep some node structs simpler, and have
   760  // higher-level nodes such as an entire rule rewrite.
   761  //
   762  // Note that ast.Expr is always used as-is; we don't declare our own expression
   763  // nodes.
   764  type (
   765  	File struct {
   766  		BodyBase // []*Func
   767  		Arch     arch
   768  		Suffix   string
   769  	}
   770  	Func struct {
   771  		BodyBase
   772  		Kind   string // "Value" or "Block"
   773  		Suffix string
   774  		ArgLen int32 // if kind == "Value", number of args for this op
   775  	}
   776  	Switch struct {
   777  		BodyBase // []*Case
   778  		Expr     ast.Expr
   779  	}
   780  	Case struct {
   781  		BodyBase
   782  		Expr ast.Expr
   783  	}
   784  	RuleRewrite struct {
   785  		BodyBase
   786  		Match, Cond, Result string // top comments
   787  		Check               string // top-level boolean expression
   788  
   789  		Alloc        int    // for unique var names
   790  		Loc          string // file name & line number of the original rule
   791  		CommuteDepth int    // used to track depth of commute loops
   792  	}
   793  	Declare struct {
   794  		Name  string
   795  		Value ast.Expr
   796  	}
   797  	CondBreak struct {
   798  		Cond              ast.Expr
   799  		InsideCommuteLoop bool
   800  	}
   801  	StartCommuteLoop struct {
   802  		Depth int
   803  		V     string
   804  	}
   805  )
   806  
   807  // exprf parses a Go expression generated from fmt.Sprintf, panicking if an
   808  // error occurs.
   809  func exprf(format string, a ...interface{}) ast.Expr {
   810  	src := fmt.Sprintf(format, a...)
   811  	expr, err := parser.ParseExpr(src)
   812  	if err != nil {
   813  		log.Fatalf("expr parse error on %q: %v", src, err)
   814  	}
   815  	return expr
   816  }
   817  
   818  // stmtf parses a Go statement generated from fmt.Sprintf. This function is only
   819  // meant for simple statements that don't have a custom Statement node declared
   820  // in this package, such as ast.ReturnStmt or ast.ExprStmt.
   821  func stmtf(format string, a ...interface{}) Statement {
   822  	src := fmt.Sprintf(format, a...)
   823  	fsrc := "package p\nfunc _() {\n" + src + "\n}\n"
   824  	file, err := parser.ParseFile(token.NewFileSet(), "", fsrc, 0)
   825  	if err != nil {
   826  		log.Fatalf("stmt parse error on %q: %v", src, err)
   827  	}
   828  	return file.Decls[0].(*ast.FuncDecl).Body.List[0]
   829  }
   830  
   831  var reservedNames = map[string]bool{
   832  	"v":      true, // Values[i], etc
   833  	"b":      true, // v.Block
   834  	"config": true, // b.Func.Config
   835  	"fe":     true, // b.Func.fe
   836  	"typ":    true, // &b.Func.Config.Types
   837  }
   838  
   839  // declf constructs a simple "name := value" declaration,
   840  // using exprf for its value.
   841  //
   842  // name must not be one of reservedNames.
   843  // This helps prevent unintended shadowing and name clashes.
   844  // To declare a reserved name, use declReserved.
   845  func declf(loc, name, format string, a ...interface{}) *Declare {
   846  	if reservedNames[name] {
   847  		log.Fatalf("rule %s uses the reserved name %s", loc, name)
   848  	}
   849  	return &Declare{name, exprf(format, a...)}
   850  }
   851  
   852  // declReserved is like declf, but the name must be one of reservedNames.
   853  // Calls to declReserved should generally be static and top-level.
   854  func declReserved(name, value string) *Declare {
   855  	if !reservedNames[name] {
   856  		panic(fmt.Sprintf("declReserved call does not use a reserved name: %q", name))
   857  	}
   858  	return &Declare{name, exprf(value)}
   859  }
   860  
   861  // breakf constructs a simple "if cond { break }" statement, using exprf for its
   862  // condition.
   863  func breakf(format string, a ...interface{}) *CondBreak {
   864  	return &CondBreak{Cond: exprf(format, a...)}
   865  }
   866  
   867  func genBlockRewrite(rule Rule, arch arch, data blockData) *RuleRewrite {
   868  	rr := &RuleRewrite{Loc: rule.Loc}
   869  	rr.Match, rr.Cond, rr.Result = rule.parse()
   870  	_, _, auxint, aux, s := extract(rr.Match) // remove parens, then split
   871  
   872  	// check match of control values
   873  	if len(s) < data.controls {
   874  		log.Fatalf("incorrect number of arguments in %s, got %v wanted at least %v", rule, len(s), data.controls)
   875  	}
   876  	controls := s[:data.controls]
   877  	pos := make([]string, data.controls)
   878  	for i, arg := range controls {
   879  		cname := fmt.Sprintf("b.Controls[%v]", i)
   880  		if strings.Contains(arg, "(") {
   881  			vname, expr := splitNameExpr(arg)
   882  			if vname == "" {
   883  				vname = fmt.Sprintf("v_%v", i)
   884  			}
   885  			rr.add(declf(rr.Loc, vname, cname))
   886  			p, op := genMatch0(rr, arch, expr, vname, nil, false) // TODO: pass non-nil cnt?
   887  			if op != "" {
   888  				check := fmt.Sprintf("%s.Op == %s", cname, op)
   889  				if rr.Check == "" {
   890  					rr.Check = check
   891  				} else {
   892  					rr.Check += " && " + check
   893  				}
   894  			}
   895  			if p == "" {
   896  				p = vname + ".Pos"
   897  			}
   898  			pos[i] = p
   899  		} else {
   900  			rr.add(declf(rr.Loc, arg, cname))
   901  			pos[i] = arg + ".Pos"
   902  		}
   903  	}
   904  	for _, e := range []struct {
   905  		name, field, dclType string
   906  	}{
   907  		{auxint, "AuxInt", data.auxIntType()},
   908  		{aux, "Aux", data.auxType()},
   909  	} {
   910  		if e.name == "" {
   911  			continue
   912  		}
   913  
   914  		if e.dclType == "" {
   915  			log.Fatalf("op %s has no declared type for %s", data.name, e.field)
   916  		}
   917  		if !token.IsIdentifier(e.name) || rr.declared(e.name) {
   918  			rr.add(breakf("%sTo%s(b.%s) != %s", unTitle(e.field), title(e.dclType), e.field, e.name))
   919  		} else {
   920  			rr.add(declf(rr.Loc, e.name, "%sTo%s(b.%s)", unTitle(e.field), title(e.dclType), e.field))
   921  		}
   922  	}
   923  	if rr.Cond != "" {
   924  		rr.add(breakf("!(%s)", rr.Cond))
   925  	}
   926  
   927  	// Rule matches. Generate result.
   928  	outop, _, auxint, aux, t := extract(rr.Result) // remove parens, then split
   929  	blockName, outdata := getBlockInfo(outop, arch)
   930  	if len(t) < outdata.controls {
   931  		log.Fatalf("incorrect number of output arguments in %s, got %v wanted at least %v", rule, len(s), outdata.controls)
   932  	}
   933  
   934  	// Check if newsuccs is the same set as succs.
   935  	succs := s[data.controls:]
   936  	newsuccs := t[outdata.controls:]
   937  	m := map[string]bool{}
   938  	for _, succ := range succs {
   939  		if m[succ] {
   940  			log.Fatalf("can't have a repeat successor name %s in %s", succ, rule)
   941  		}
   942  		m[succ] = true
   943  	}
   944  	for _, succ := range newsuccs {
   945  		if !m[succ] {
   946  			log.Fatalf("unknown successor %s in %s", succ, rule)
   947  		}
   948  		delete(m, succ)
   949  	}
   950  	if len(m) != 0 {
   951  		log.Fatalf("unmatched successors %v in %s", m, rule)
   952  	}
   953  
   954  	var genControls [2]string
   955  	for i, control := range t[:outdata.controls] {
   956  		// Select a source position for any new control values.
   957  		// TODO: does it always make sense to use the source position
   958  		// of the original control values or should we be using the
   959  		// block's source position in some cases?
   960  		newpos := "b.Pos" // default to block's source position
   961  		if i < len(pos) && pos[i] != "" {
   962  			// Use the previous control value's source position.
   963  			newpos = pos[i]
   964  		}
   965  
   966  		// Generate a new control value (or copy an existing value).
   967  		genControls[i] = genResult0(rr, arch, control, false, false, newpos, nil)
   968  	}
   969  	switch outdata.controls {
   970  	case 0:
   971  		rr.add(stmtf("b.Reset(%s)", blockName))
   972  	case 1:
   973  		rr.add(stmtf("b.resetWithControl(%s, %s)", blockName, genControls[0]))
   974  	case 2:
   975  		rr.add(stmtf("b.resetWithControl2(%s, %s, %s)", blockName, genControls[0], genControls[1]))
   976  	default:
   977  		log.Fatalf("too many controls: %d", outdata.controls)
   978  	}
   979  
   980  	if auxint != "" {
   981  		// Make sure auxint value has the right type.
   982  		rr.add(stmtf("b.AuxInt = %sToAuxInt(%s)", unTitle(outdata.auxIntType()), auxint))
   983  	}
   984  	if aux != "" {
   985  		// Make sure aux value has the right type.
   986  		rr.add(stmtf("b.Aux = %sToAux(%s)", unTitle(outdata.auxType()), aux))
   987  	}
   988  
   989  	succChanged := false
   990  	for i := 0; i < len(succs); i++ {
   991  		if succs[i] != newsuccs[i] {
   992  			succChanged = true
   993  		}
   994  	}
   995  	if succChanged {
   996  		if len(succs) != 2 {
   997  			log.Fatalf("changed successors, len!=2 in %s", rule)
   998  		}
   999  		if succs[0] != newsuccs[1] || succs[1] != newsuccs[0] {
  1000  			log.Fatalf("can only handle swapped successors in %s", rule)
  1001  		}
  1002  		rr.add(stmtf("b.swapSuccessors()"))
  1003  	}
  1004  
  1005  	if *genLog {
  1006  		rr.add(stmtf("logRule(%q)", rule.Loc))
  1007  	}
  1008  	return rr
  1009  }
  1010  
  1011  // genMatch returns the variable whose source position should be used for the
  1012  // result (or "" if no opinion), and a boolean that reports whether the match can fail.
  1013  func genMatch(rr *RuleRewrite, arch arch, match string, pregenTop bool) (pos, checkOp string) {
  1014  	cnt := varCount(rr)
  1015  	return genMatch0(rr, arch, match, "v", cnt, pregenTop)
  1016  }
  1017  
  1018  func genMatch0(rr *RuleRewrite, arch arch, match, v string, cnt map[string]int, pregenTop bool) (pos, checkOp string) {
  1019  	if match[0] != '(' || match[len(match)-1] != ')' {
  1020  		log.Fatalf("%s: non-compound expr in genMatch0: %q", rr.Loc, match)
  1021  	}
  1022  	op, oparch, typ, auxint, aux, args := parseValue(match, arch, rr.Loc)
  1023  
  1024  	checkOp = fmt.Sprintf("Op%s%s", oparch, op.name)
  1025  
  1026  	if op.faultOnNilArg0 || op.faultOnNilArg1 {
  1027  		// Prefer the position of an instruction which could fault.
  1028  		pos = v + ".Pos"
  1029  	}
  1030  
  1031  	// If the last argument is ___, it means "don't care about trailing arguments, really"
  1032  	// The likely/intended use is for rewrites that are too tricky to express in the existing pattern language
  1033  	// Do a length check early because long patterns fed short (ultimately not-matching) inputs will
  1034  	// do an indexing error in pattern-matching.
  1035  	if op.argLength == -1 {
  1036  		l := len(args)
  1037  		if l == 0 || args[l-1] != "___" {
  1038  			rr.add(breakf("len(%s.Args) != %d", v, l))
  1039  		} else if l > 1 && args[l-1] == "___" {
  1040  			rr.add(breakf("len(%s.Args) < %d", v, l-1))
  1041  		}
  1042  	}
  1043  
  1044  	for _, e := range []struct {
  1045  		name, field, dclType string
  1046  	}{
  1047  		{typ, "Type", "*types.Type"},
  1048  		{auxint, "AuxInt", op.auxIntType()},
  1049  		{aux, "Aux", op.auxType()},
  1050  	} {
  1051  		if e.name == "" {
  1052  			continue
  1053  		}
  1054  
  1055  		if e.dclType == "" {
  1056  			log.Fatalf("op %s has no declared type for %s", op.name, e.field)
  1057  		}
  1058  		if !token.IsIdentifier(e.name) || rr.declared(e.name) {
  1059  			switch e.field {
  1060  			case "Aux":
  1061  				rr.add(breakf("auxTo%s(%s.%s) != %s", title(e.dclType), v, e.field, e.name))
  1062  			case "AuxInt":
  1063  				rr.add(breakf("auxIntTo%s(%s.%s) != %s", title(e.dclType), v, e.field, e.name))
  1064  			case "Type":
  1065  				rr.add(breakf("%s.%s != %s", v, e.field, e.name))
  1066  			}
  1067  		} else {
  1068  			switch e.field {
  1069  			case "Aux":
  1070  				rr.add(declf(rr.Loc, e.name, "auxTo%s(%s.%s)", title(e.dclType), v, e.field))
  1071  			case "AuxInt":
  1072  				rr.add(declf(rr.Loc, e.name, "auxIntTo%s(%s.%s)", title(e.dclType), v, e.field))
  1073  			case "Type":
  1074  				rr.add(declf(rr.Loc, e.name, "%s.%s", v, e.field))
  1075  			}
  1076  		}
  1077  	}
  1078  
  1079  	commutative := op.commutative
  1080  	if commutative {
  1081  		if args[0] == args[1] {
  1082  			// When we have (Add x x), for any x,
  1083  			// even if there are other uses of x besides these two,
  1084  			// and even if x is not a variable,
  1085  			// we can skip the commutative match.
  1086  			commutative = false
  1087  		}
  1088  		if cnt[args[0]] == 1 && cnt[args[1]] == 1 {
  1089  			// When we have (Add x y) with no other uses
  1090  			// of x and y in the matching rule and condition,
  1091  			// then we can skip the commutative match (Add y x).
  1092  			commutative = false
  1093  		}
  1094  	}
  1095  
  1096  	if !pregenTop {
  1097  		// Access last argument first to minimize bounds checks.
  1098  		for n := len(args) - 1; n > 0; n-- {
  1099  			a := args[n]
  1100  			if a == "_" {
  1101  				continue
  1102  			}
  1103  			if !rr.declared(a) && token.IsIdentifier(a) && !(commutative && len(args) == 2) {
  1104  				rr.add(declf(rr.Loc, a, "%s.Args[%d]", v, n))
  1105  				// delete the last argument so it is not reprocessed
  1106  				args = args[:n]
  1107  			} else {
  1108  				rr.add(stmtf("_ = %s.Args[%d]", v, n))
  1109  			}
  1110  			break
  1111  		}
  1112  	}
  1113  	if commutative && !pregenTop {
  1114  		for i := 0; i <= 1; i++ {
  1115  			vname := fmt.Sprintf("%s_%d", v, i)
  1116  			rr.add(declf(rr.Loc, vname, "%s.Args[%d]", v, i))
  1117  		}
  1118  	}
  1119  	if commutative {
  1120  		rr.add(StartCommuteLoop{rr.CommuteDepth, v})
  1121  		rr.CommuteDepth++
  1122  	}
  1123  	for i, arg := range args {
  1124  		if arg == "_" {
  1125  			continue
  1126  		}
  1127  		var rhs string
  1128  		if (commutative && i < 2) || pregenTop {
  1129  			rhs = fmt.Sprintf("%s_%d", v, i)
  1130  		} else {
  1131  			rhs = fmt.Sprintf("%s.Args[%d]", v, i)
  1132  		}
  1133  		if !strings.Contains(arg, "(") {
  1134  			// leaf variable
  1135  			if rr.declared(arg) {
  1136  				// variable already has a definition. Check whether
  1137  				// the old definition and the new definition match.
  1138  				// For example, (add x x).  Equality is just pointer equality
  1139  				// on Values (so cse is important to do before lowering).
  1140  				rr.add(breakf("%s != %s", arg, rhs))
  1141  			} else {
  1142  				if arg != rhs {
  1143  					rr.add(declf(rr.Loc, arg, "%s", rhs))
  1144  				}
  1145  			}
  1146  			continue
  1147  		}
  1148  		// compound sexpr
  1149  		argname, expr := splitNameExpr(arg)
  1150  		if argname == "" {
  1151  			argname = fmt.Sprintf("%s_%d", v, i)
  1152  		}
  1153  		if argname == "b" {
  1154  			log.Fatalf("don't name args 'b', it is ambiguous with blocks")
  1155  		}
  1156  
  1157  		if argname != rhs {
  1158  			rr.add(declf(rr.Loc, argname, "%s", rhs))
  1159  		}
  1160  		bexpr := exprf("%s.Op != addLater", argname)
  1161  		rr.add(&CondBreak{Cond: bexpr})
  1162  		argPos, argCheckOp := genMatch0(rr, arch, expr, argname, cnt, false)
  1163  		bexpr.(*ast.BinaryExpr).Y.(*ast.Ident).Name = argCheckOp
  1164  
  1165  		if argPos != "" {
  1166  			// Keep the argument in preference to the parent, as the
  1167  			// argument is normally earlier in program flow.
  1168  			// Keep the argument in preference to an earlier argument,
  1169  			// as that prefers the memory argument which is also earlier
  1170  			// in the program flow.
  1171  			pos = argPos
  1172  		}
  1173  	}
  1174  
  1175  	return pos, checkOp
  1176  }
  1177  
  1178  func genResult(rr *RuleRewrite, arch arch, result, pos string) {
  1179  	move := result[0] == '@'
  1180  	if move {
  1181  		// parse @block directive
  1182  		s := strings.SplitN(result[1:], " ", 2)
  1183  		rr.add(stmtf("b = %s", s[0]))
  1184  		result = s[1]
  1185  	}
  1186  	if result[0] == '{' {
  1187  		// Arbitrary code used to make the result
  1188  		rr.add(stmtf("v.copyOf(%s)", result[1:len(result)-1]))
  1189  		return
  1190  	}
  1191  	cse := make(map[string]string)
  1192  	genResult0(rr, arch, result, true, move, pos, cse)
  1193  }
  1194  
  1195  func genResult0(rr *RuleRewrite, arch arch, result string, top, move bool, pos string, cse map[string]string) string {
  1196  	resname, expr := splitNameExpr(result)
  1197  	result = expr
  1198  	// TODO: when generating a constant result, use f.constVal to avoid
  1199  	// introducing copies just to clean them up again.
  1200  	if result[0] != '(' {
  1201  		// variable
  1202  		if top {
  1203  			// It in not safe in general to move a variable between blocks
  1204  			// (and particularly not a phi node).
  1205  			// Introduce a copy.
  1206  			rr.add(stmtf("v.copyOf(%s)", result))
  1207  		}
  1208  		return result
  1209  	}
  1210  
  1211  	w := normalizeWhitespace(result)
  1212  	if prev := cse[w]; prev != "" {
  1213  		return prev
  1214  	}
  1215  
  1216  	op, oparch, typ, auxint, aux, args := parseValue(result, arch, rr.Loc)
  1217  
  1218  	// Find the type of the variable.
  1219  	typeOverride := typ != ""
  1220  	if typ == "" && op.typ != "" {
  1221  		typ = typeName(op.typ)
  1222  	}
  1223  
  1224  	v := "v"
  1225  	if top && !move {
  1226  		rr.add(stmtf("v.reset(Op%s%s)", oparch, op.name))
  1227  		if typeOverride {
  1228  			rr.add(stmtf("v.Type = %s", typ))
  1229  		}
  1230  	} else {
  1231  		if typ == "" {
  1232  			log.Fatalf("sub-expression %s (op=Op%s%s) at %s must have a type", result, oparch, op.name, rr.Loc)
  1233  		}
  1234  		if resname == "" {
  1235  			v = fmt.Sprintf("v%d", rr.Alloc)
  1236  		} else {
  1237  			v = resname
  1238  		}
  1239  		rr.Alloc++
  1240  		rr.add(declf(rr.Loc, v, "b.NewValue0(%s, Op%s%s, %s)", pos, oparch, op.name, typ))
  1241  		if move && top {
  1242  			// Rewrite original into a copy
  1243  			rr.add(stmtf("v.copyOf(%s)", v))
  1244  		}
  1245  	}
  1246  
  1247  	if auxint != "" {
  1248  		// Make sure auxint value has the right type.
  1249  		rr.add(stmtf("%s.AuxInt = %sToAuxInt(%s)", v, unTitle(op.auxIntType()), auxint))
  1250  	}
  1251  	if aux != "" {
  1252  		// Make sure aux value has the right type.
  1253  		rr.add(stmtf("%s.Aux = %sToAux(%s)", v, unTitle(op.auxType()), aux))
  1254  	}
  1255  	all := new(strings.Builder)
  1256  	for i, arg := range args {
  1257  		x := genResult0(rr, arch, arg, false, move, pos, cse)
  1258  		if i > 0 {
  1259  			all.WriteString(", ")
  1260  		}
  1261  		all.WriteString(x)
  1262  	}
  1263  	switch len(args) {
  1264  	case 0:
  1265  	case 1:
  1266  		rr.add(stmtf("%s.AddArg(%s)", v, all.String()))
  1267  	default:
  1268  		rr.add(stmtf("%s.AddArg%d(%s)", v, len(args), all.String()))
  1269  	}
  1270  
  1271  	if cse != nil {
  1272  		cse[w] = v
  1273  	}
  1274  	return v
  1275  }
  1276  
  1277  func split(s string) []string {
  1278  	var r []string
  1279  
  1280  outer:
  1281  	for s != "" {
  1282  		d := 0               // depth of ({[<
  1283  		var open, close byte // opening and closing markers ({[< or )}]>
  1284  		nonsp := false       // found a non-space char so far
  1285  		for i := 0; i < len(s); i++ {
  1286  			switch {
  1287  			case d == 0 && s[i] == '(':
  1288  				open, close = '(', ')'
  1289  				d++
  1290  			case d == 0 && s[i] == '<':
  1291  				open, close = '<', '>'
  1292  				d++
  1293  			case d == 0 && s[i] == '[':
  1294  				open, close = '[', ']'
  1295  				d++
  1296  			case d == 0 && s[i] == '{':
  1297  				open, close = '{', '}'
  1298  				d++
  1299  			case d == 0 && (s[i] == ' ' || s[i] == '\t'):
  1300  				if nonsp {
  1301  					r = append(r, strings.TrimSpace(s[:i]))
  1302  					s = s[i:]
  1303  					continue outer
  1304  				}
  1305  			case d > 0 && s[i] == open:
  1306  				d++
  1307  			case d > 0 && s[i] == close:
  1308  				d--
  1309  			default:
  1310  				nonsp = true
  1311  			}
  1312  		}
  1313  		if d != 0 {
  1314  			log.Fatalf("imbalanced expression: %q", s)
  1315  		}
  1316  		if nonsp {
  1317  			r = append(r, strings.TrimSpace(s))
  1318  		}
  1319  		break
  1320  	}
  1321  	return r
  1322  }
  1323  
  1324  // isBlock reports whether this op is a block opcode.
  1325  func isBlock(name string, arch arch) bool {
  1326  	for _, b := range genericBlocks {
  1327  		if b.name == name {
  1328  			return true
  1329  		}
  1330  	}
  1331  	for _, b := range arch.blocks {
  1332  		if b.name == name {
  1333  			return true
  1334  		}
  1335  	}
  1336  	return false
  1337  }
  1338  
  1339  func extract(val string) (op, typ, auxint, aux string, args []string) {
  1340  	val = val[1 : len(val)-1] // remove ()
  1341  
  1342  	// Split val up into regions.
  1343  	// Split by spaces/tabs, except those contained in (), {}, [], or <>.
  1344  	s := split(val)
  1345  
  1346  	// Extract restrictions and args.
  1347  	op = s[0]
  1348  	for _, a := range s[1:] {
  1349  		switch a[0] {
  1350  		case '<':
  1351  			typ = a[1 : len(a)-1] // remove <>
  1352  		case '[':
  1353  			auxint = a[1 : len(a)-1] // remove []
  1354  		case '{':
  1355  			aux = a[1 : len(a)-1] // remove {}
  1356  		default:
  1357  			args = append(args, a)
  1358  		}
  1359  	}
  1360  	return
  1361  }
  1362  
  1363  // parseValue parses a parenthesized value from a rule.
  1364  // The value can be from the match or the result side.
  1365  // It returns the op and unparsed strings for typ, auxint, and aux restrictions and for all args.
  1366  // oparch is the architecture that op is located in, or "" for generic.
  1367  func parseValue(val string, arch arch, loc string) (op opData, oparch, typ, auxint, aux string, args []string) {
  1368  	// Resolve the op.
  1369  	var s string
  1370  	s, typ, auxint, aux, args = extract(val)
  1371  
  1372  	// match reports whether x is a good op to select.
  1373  	// If strict is true, rule generation might succeed.
  1374  	// If strict is false, rule generation has failed,
  1375  	// but we're trying to generate a useful error.
  1376  	// Doing strict=true then strict=false allows
  1377  	// precise op matching while retaining good error messages.
  1378  	match := func(x opData, strict bool, archname string) bool {
  1379  		if x.name != s {
  1380  			return false
  1381  		}
  1382  		if x.argLength != -1 && int(x.argLength) != len(args) && (len(args) != 1 || args[0] != "...") {
  1383  			if strict {
  1384  				return false
  1385  			}
  1386  			log.Printf("%s: op %s (%s) should have %d args, has %d", loc, s, archname, x.argLength, len(args))
  1387  		}
  1388  		return true
  1389  	}
  1390  
  1391  	for _, x := range genericOps {
  1392  		if match(x, true, "generic") {
  1393  			op = x
  1394  			break
  1395  		}
  1396  	}
  1397  	for _, x := range arch.ops {
  1398  		if arch.name != "generic" && match(x, true, arch.name) {
  1399  			if op.name != "" {
  1400  				log.Fatalf("%s: matches for op %s found in both generic and %s", loc, op.name, arch.name)
  1401  			}
  1402  			op = x
  1403  			oparch = arch.name
  1404  			break
  1405  		}
  1406  	}
  1407  
  1408  	if op.name == "" {
  1409  		// Failed to find the op.
  1410  		// Run through everything again with strict=false
  1411  		// to generate useful diagnostic messages before failing.
  1412  		for _, x := range genericOps {
  1413  			match(x, false, "generic")
  1414  		}
  1415  		for _, x := range arch.ops {
  1416  			match(x, false, arch.name)
  1417  		}
  1418  		log.Fatalf("%s: unknown op %s", loc, s)
  1419  	}
  1420  
  1421  	// Sanity check aux, auxint.
  1422  	if auxint != "" && !opHasAuxInt(op) {
  1423  		log.Fatalf("%s: op %s %s can't have auxint", loc, op.name, op.aux)
  1424  	}
  1425  	if aux != "" && !opHasAux(op) {
  1426  		log.Fatalf("%s: op %s %s can't have aux", loc, op.name, op.aux)
  1427  	}
  1428  	return
  1429  }
  1430  
  1431  func opHasAuxInt(op opData) bool {
  1432  	switch op.aux {
  1433  	case "Bool", "Int8", "Int16", "Int32", "Int64", "Int128", "UInt8", "Float32", "Float64",
  1434  		"SymOff", "CallOff", "SymValAndOff", "TypSize", "ARM64BitField", "FlagConstant", "CCop":
  1435  		return true
  1436  	}
  1437  	return false
  1438  }
  1439  
  1440  func opHasAux(op opData) bool {
  1441  	switch op.aux {
  1442  	case "String", "Sym", "SymOff", "Call", "CallOff", "SymValAndOff", "Typ", "TypSize",
  1443  		"S390XCCMask", "S390XRotateParams":
  1444  		return true
  1445  	}
  1446  	return false
  1447  }
  1448  
  1449  // splitNameExpr splits s-expr arg, possibly prefixed by "name:",
  1450  // into name and the unprefixed expression.
  1451  // For example, "x:(Foo)" yields "x", "(Foo)",
  1452  // and "(Foo)" yields "", "(Foo)".
  1453  func splitNameExpr(arg string) (name, expr string) {
  1454  	colon := strings.Index(arg, ":")
  1455  	if colon < 0 {
  1456  		return "", arg
  1457  	}
  1458  	openparen := strings.Index(arg, "(")
  1459  	if openparen < 0 {
  1460  		log.Fatalf("splitNameExpr(%q): colon but no open parens", arg)
  1461  	}
  1462  	if colon > openparen {
  1463  		// colon is inside the parens, such as in "(Foo x:(Bar))".
  1464  		return "", arg
  1465  	}
  1466  	return arg[:colon], arg[colon+1:]
  1467  }
  1468  
  1469  func getBlockInfo(op string, arch arch) (name string, data blockData) {
  1470  	for _, b := range genericBlocks {
  1471  		if b.name == op {
  1472  			return "Block" + op, b
  1473  		}
  1474  	}
  1475  	for _, b := range arch.blocks {
  1476  		if b.name == op {
  1477  			return "Block" + arch.name + op, b
  1478  		}
  1479  	}
  1480  	log.Fatalf("could not find block data for %s", op)
  1481  	panic("unreachable")
  1482  }
  1483  
  1484  // typeName returns the string to use to generate a type.
  1485  func typeName(typ string) string {
  1486  	if typ[0] == '(' {
  1487  		ts := strings.Split(typ[1:len(typ)-1], ",")
  1488  		if len(ts) != 2 {
  1489  			log.Fatalf("Tuple expect 2 arguments")
  1490  		}
  1491  		return "types.NewTuple(" + typeName(ts[0]) + ", " + typeName(ts[1]) + ")"
  1492  	}
  1493  	switch typ {
  1494  	case "Flags", "Mem", "Void", "Int128":
  1495  		return "types.Type" + typ
  1496  	default:
  1497  		return "typ." + typ
  1498  	}
  1499  }
  1500  
  1501  // balance returns the number of unclosed '(' characters in s.
  1502  // If a ')' appears without a corresponding '(', balance returns -1.
  1503  func balance(s string) int {
  1504  	balance := 0
  1505  	for _, c := range s {
  1506  		switch c {
  1507  		case '(':
  1508  			balance++
  1509  		case ')':
  1510  			balance--
  1511  			if balance < 0 {
  1512  				// don't allow ")(" to return 0
  1513  				return -1
  1514  			}
  1515  		}
  1516  	}
  1517  	return balance
  1518  }
  1519  
  1520  // findAllOpcode is a function to find the opcode portion of s-expressions.
  1521  var findAllOpcode = regexp.MustCompile(`[(](\w+[|])+\w+[)]`).FindAllStringIndex
  1522  
  1523  // excludeFromExpansion reports whether the substring s[idx[0]:idx[1]] in a rule
  1524  // should be disregarded as a candidate for | expansion.
  1525  // It uses simple syntactic checks to see whether the substring
  1526  // is inside an AuxInt expression or inside the && conditions.
  1527  func excludeFromExpansion(s string, idx []int) bool {
  1528  	left := s[:idx[0]]
  1529  	if strings.LastIndexByte(left, '[') > strings.LastIndexByte(left, ']') {
  1530  		// Inside an AuxInt expression.
  1531  		return true
  1532  	}
  1533  	right := s[idx[1]:]
  1534  	if strings.Contains(left, "&&") && strings.Contains(right, "=>") {
  1535  		// Inside && conditions.
  1536  		return true
  1537  	}
  1538  	return false
  1539  }
  1540  
  1541  // expandOr converts a rule into multiple rules by expanding | ops.
  1542  func expandOr(r string) []string {
  1543  	// Find every occurrence of |-separated things.
  1544  	// They look like MOV(B|W|L|Q|SS|SD)load or MOV(Q|L)loadidx(1|8).
  1545  	// Generate rules selecting one case from each |-form.
  1546  
  1547  	// Count width of |-forms.  They must match.
  1548  	n := 1
  1549  	for _, idx := range findAllOpcode(r, -1) {
  1550  		if excludeFromExpansion(r, idx) {
  1551  			continue
  1552  		}
  1553  		s := r[idx[0]:idx[1]]
  1554  		c := strings.Count(s, "|") + 1
  1555  		if c == 1 {
  1556  			continue
  1557  		}
  1558  		if n > 1 && n != c {
  1559  			log.Fatalf("'|' count doesn't match in %s: both %d and %d\n", r, n, c)
  1560  		}
  1561  		n = c
  1562  	}
  1563  	if n == 1 {
  1564  		// No |-form in this rule.
  1565  		return []string{r}
  1566  	}
  1567  	// Build each new rule.
  1568  	res := make([]string, n)
  1569  	for i := 0; i < n; i++ {
  1570  		buf := new(strings.Builder)
  1571  		x := 0
  1572  		for _, idx := range findAllOpcode(r, -1) {
  1573  			if excludeFromExpansion(r, idx) {
  1574  				continue
  1575  			}
  1576  			buf.WriteString(r[x:idx[0]])              // write bytes we've skipped over so far
  1577  			s := r[idx[0]+1 : idx[1]-1]               // remove leading "(" and trailing ")"
  1578  			buf.WriteString(strings.Split(s, "|")[i]) // write the op component for this rule
  1579  			x = idx[1]                                // note that we've written more bytes
  1580  		}
  1581  		buf.WriteString(r[x:])
  1582  		res[i] = buf.String()
  1583  	}
  1584  	return res
  1585  }
  1586  
  1587  // varCount returns a map which counts the number of occurrences of
  1588  // Value variables in the s-expression rr.Match and the Go expression rr.Cond.
  1589  func varCount(rr *RuleRewrite) map[string]int {
  1590  	cnt := map[string]int{}
  1591  	varCount1(rr.Loc, rr.Match, cnt)
  1592  	if rr.Cond != "" {
  1593  		expr, err := parser.ParseExpr(rr.Cond)
  1594  		if err != nil {
  1595  			log.Fatalf("%s: failed to parse cond %q: %v", rr.Loc, rr.Cond, err)
  1596  		}
  1597  		ast.Inspect(expr, func(n ast.Node) bool {
  1598  			if id, ok := n.(*ast.Ident); ok {
  1599  				cnt[id.Name]++
  1600  			}
  1601  			return true
  1602  		})
  1603  	}
  1604  	return cnt
  1605  }
  1606  
  1607  func varCount1(loc, m string, cnt map[string]int) {
  1608  	if m[0] == '<' || m[0] == '[' || m[0] == '{' {
  1609  		return
  1610  	}
  1611  	if token.IsIdentifier(m) {
  1612  		cnt[m]++
  1613  		return
  1614  	}
  1615  	// Split up input.
  1616  	name, expr := splitNameExpr(m)
  1617  	if name != "" {
  1618  		cnt[name]++
  1619  	}
  1620  	if expr[0] != '(' || expr[len(expr)-1] != ')' {
  1621  		log.Fatalf("%s: non-compound expr in varCount1: %q", loc, expr)
  1622  	}
  1623  	s := split(expr[1 : len(expr)-1])
  1624  	for _, arg := range s[1:] {
  1625  		varCount1(loc, arg, cnt)
  1626  	}
  1627  }
  1628  
  1629  // normalizeWhitespace replaces 2+ whitespace sequences with a single space.
  1630  func normalizeWhitespace(x string) string {
  1631  	x = strings.Join(strings.Fields(x), " ")
  1632  	x = strings.ReplaceAll(x, "( ", "(")
  1633  	x = strings.ReplaceAll(x, " )", ")")
  1634  	x = strings.ReplaceAll(x, "[ ", "[")
  1635  	x = strings.ReplaceAll(x, " ]", "]")
  1636  	x = strings.ReplaceAll(x, ")=>", ") =>")
  1637  	return x
  1638  }
  1639  
  1640  // opIsCommutative reports whether op s is commutative.
  1641  func opIsCommutative(op string, arch arch) bool {
  1642  	for _, x := range genericOps {
  1643  		if op == x.name {
  1644  			if x.commutative {
  1645  				return true
  1646  			}
  1647  			break
  1648  		}
  1649  	}
  1650  	if arch.name != "generic" {
  1651  		for _, x := range arch.ops {
  1652  			if op == x.name {
  1653  				if x.commutative {
  1654  					return true
  1655  				}
  1656  				break
  1657  			}
  1658  		}
  1659  	}
  1660  	return false
  1661  }
  1662  
  1663  func normalizeMatch(m string, arch arch) string {
  1664  	if token.IsIdentifier(m) {
  1665  		return m
  1666  	}
  1667  	op, typ, auxint, aux, args := extract(m)
  1668  	if opIsCommutative(op, arch) {
  1669  		if args[1] < args[0] {
  1670  			args[0], args[1] = args[1], args[0]
  1671  		}
  1672  	}
  1673  	s := new(strings.Builder)
  1674  	fmt.Fprintf(s, "%s <%s> [%s] {%s}", op, typ, auxint, aux)
  1675  	for _, arg := range args {
  1676  		prefix, expr := splitNameExpr(arg)
  1677  		fmt.Fprint(s, " ", prefix, normalizeMatch(expr, arch))
  1678  	}
  1679  	return s.String()
  1680  }
  1681  
  1682  func parseEllipsisRules(rules []Rule, arch arch) (newop string, ok bool) {
  1683  	if len(rules) != 1 {
  1684  		for _, r := range rules {
  1685  			if strings.Contains(r.Rule, "...") {
  1686  				log.Fatalf("%s: found ellipsis in rule, but there are other rules with the same op", r.Loc)
  1687  			}
  1688  		}
  1689  		return "", false
  1690  	}
  1691  	rule := rules[0]
  1692  	match, cond, result := rule.parse()
  1693  	if cond != "" || !isEllipsisValue(match) || !isEllipsisValue(result) {
  1694  		if strings.Contains(rule.Rule, "...") {
  1695  			log.Fatalf("%s: found ellipsis in non-ellipsis rule", rule.Loc)
  1696  		}
  1697  		checkEllipsisRuleCandidate(rule, arch)
  1698  		return "", false
  1699  	}
  1700  	op, oparch, _, _, _, _ := parseValue(result, arch, rule.Loc)
  1701  	return fmt.Sprintf("Op%s%s", oparch, op.name), true
  1702  }
  1703  
  1704  // isEllipsisValue reports whether s is of the form (OpX ...).
  1705  func isEllipsisValue(s string) bool {
  1706  	if len(s) < 2 || s[0] != '(' || s[len(s)-1] != ')' {
  1707  		return false
  1708  	}
  1709  	c := split(s[1 : len(s)-1])
  1710  	if len(c) != 2 || c[1] != "..." {
  1711  		return false
  1712  	}
  1713  	return true
  1714  }
  1715  
  1716  func checkEllipsisRuleCandidate(rule Rule, arch arch) {
  1717  	match, cond, result := rule.parse()
  1718  	if cond != "" {
  1719  		return
  1720  	}
  1721  	op, _, _, auxint, aux, args := parseValue(match, arch, rule.Loc)
  1722  	var auxint2, aux2 string
  1723  	var args2 []string
  1724  	var usingCopy string
  1725  	var eop opData
  1726  	if result[0] != '(' {
  1727  		// Check for (Foo x) => x, which can be converted to (Foo ...) => (Copy ...).
  1728  		args2 = []string{result}
  1729  		usingCopy = " using Copy"
  1730  	} else {
  1731  		eop, _, _, auxint2, aux2, args2 = parseValue(result, arch, rule.Loc)
  1732  	}
  1733  	// Check that all restrictions in match are reproduced exactly in result.
  1734  	if aux != aux2 || auxint != auxint2 || len(args) != len(args2) {
  1735  		return
  1736  	}
  1737  	if strings.Contains(rule.Rule, "=>") && op.aux != eop.aux {
  1738  		return
  1739  	}
  1740  	for i := range args {
  1741  		if args[i] != args2[i] {
  1742  			return
  1743  		}
  1744  	}
  1745  	switch {
  1746  	case opHasAux(op) && aux == "" && aux2 == "":
  1747  		fmt.Printf("%s: rule silently zeros aux, either copy aux or explicitly zero\n", rule.Loc)
  1748  	case opHasAuxInt(op) && auxint == "" && auxint2 == "":
  1749  		fmt.Printf("%s: rule silently zeros auxint, either copy auxint or explicitly zero\n", rule.Loc)
  1750  	default:
  1751  		fmt.Printf("%s: possible ellipsis rule candidate%s: %q\n", rule.Loc, usingCopy, rule.Rule)
  1752  	}
  1753  }
  1754  
  1755  func opByName(arch arch, name string) opData {
  1756  	name = name[2:]
  1757  	for _, x := range genericOps {
  1758  		if name == x.name {
  1759  			return x
  1760  		}
  1761  	}
  1762  	if arch.name != "generic" {
  1763  		name = name[len(arch.name):]
  1764  		for _, x := range arch.ops {
  1765  			if name == x.name {
  1766  				return x
  1767  			}
  1768  		}
  1769  	}
  1770  	log.Fatalf("failed to find op named %s in arch %s", name, arch.name)
  1771  	panic("unreachable")
  1772  }
  1773  
  1774  // auxType returns the Go type that this operation should store in its aux field.
  1775  func (op opData) auxType() string {
  1776  	switch op.aux {
  1777  	case "String":
  1778  		return "string"
  1779  	case "Sym":
  1780  		// Note: a Sym can be an *obj.LSym, a *ir.Name, or nil.
  1781  		return "Sym"
  1782  	case "SymOff":
  1783  		return "Sym"
  1784  	case "Call":
  1785  		return "Call"
  1786  	case "CallOff":
  1787  		return "Call"
  1788  	case "SymValAndOff":
  1789  		return "Sym"
  1790  	case "Typ":
  1791  		return "*types.Type"
  1792  	case "TypSize":
  1793  		return "*types.Type"
  1794  	case "S390XCCMask":
  1795  		return "s390x.CCMask"
  1796  	case "S390XRotateParams":
  1797  		return "s390x.RotateParams"
  1798  	default:
  1799  		return "invalid"
  1800  	}
  1801  }
  1802  
  1803  // auxIntType returns the Go type that this operation should store in its auxInt field.
  1804  func (op opData) auxIntType() string {
  1805  	switch op.aux {
  1806  	case "Bool":
  1807  		return "bool"
  1808  	case "Int8":
  1809  		return "int8"
  1810  	case "Int16":
  1811  		return "int16"
  1812  	case "Int32":
  1813  		return "int32"
  1814  	case "Int64":
  1815  		return "int64"
  1816  	case "Int128":
  1817  		return "int128"
  1818  	case "UInt8":
  1819  		return "uint8"
  1820  	case "Float32":
  1821  		return "float32"
  1822  	case "Float64":
  1823  		return "float64"
  1824  	case "CallOff":
  1825  		return "int32"
  1826  	case "SymOff":
  1827  		return "int32"
  1828  	case "SymValAndOff":
  1829  		return "ValAndOff"
  1830  	case "TypSize":
  1831  		return "int64"
  1832  	case "CCop":
  1833  		return "Op"
  1834  	case "FlagConstant":
  1835  		return "flagConstant"
  1836  	case "ARM64BitField":
  1837  		return "arm64BitField"
  1838  	default:
  1839  		return "invalid"
  1840  	}
  1841  }
  1842  
  1843  // auxType returns the Go type that this block should store in its aux field.
  1844  func (b blockData) auxType() string {
  1845  	switch b.aux {
  1846  	case "Sym":
  1847  		return "Sym"
  1848  	case "S390XCCMask", "S390XCCMaskInt8", "S390XCCMaskUint8":
  1849  		return "s390x.CCMask"
  1850  	case "S390XRotateParams":
  1851  		return "s390x.RotateParams"
  1852  	default:
  1853  		return "invalid"
  1854  	}
  1855  }
  1856  
  1857  // auxIntType returns the Go type that this block should store in its auxInt field.
  1858  func (b blockData) auxIntType() string {
  1859  	switch b.aux {
  1860  	case "S390XCCMaskInt8":
  1861  		return "int8"
  1862  	case "S390XCCMaskUint8":
  1863  		return "uint8"
  1864  	case "Int64":
  1865  		return "int64"
  1866  	default:
  1867  		return "invalid"
  1868  	}
  1869  }
  1870  
  1871  func title(s string) string {
  1872  	if i := strings.Index(s, "."); i >= 0 {
  1873  		switch strings.ToLower(s[:i]) {
  1874  		case "s390x": // keep arch prefix for clarity
  1875  			s = s[:i] + s[i+1:]
  1876  		default:
  1877  			s = s[i+1:]
  1878  		}
  1879  	}
  1880  	return strings.Title(s)
  1881  }
  1882  
  1883  func unTitle(s string) string {
  1884  	if i := strings.Index(s, "."); i >= 0 {
  1885  		switch strings.ToLower(s[:i]) {
  1886  		case "s390x": // keep arch prefix for clarity
  1887  			s = s[:i] + s[i+1:]
  1888  		default:
  1889  			s = s[i+1:]
  1890  		}
  1891  	}
  1892  	return strings.ToLower(s[:1]) + s[1:]
  1893  }
  1894  

View as plain text