hashdebug.go

     1  // Copyright 2022 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 base
     6  
     7  import (
     8  	"bytes"
     9  	"cmd/internal/obj"
    10  	"cmd/internal/src"
    11  	"fmt"
    12  	"internal/bisect"
    13  	"io"
    14  	"os"
    15  	"path/filepath"
    16  	"strconv"
    17  	"strings"
    18  	"sync"
    19  )
    20  
    21  type hashAndMask struct {
    22  	// a hash h matches if (h^hash)&mask == 0
    23  	hash uint64
    24  	mask uint64
    25  	name string // base name, or base name + "0", "1", etc.
    26  }
    27  
    28  type HashDebug struct {
    29  	mu   sync.Mutex // for logfile, posTmp, bytesTmp
    30  	name string     // base name of the flag/variable.
    31  	// what file (if any) receives the yes/no logging?
    32  	// default is os.Stdout
    33  	logfile          io.Writer
    34  	posTmp           []src.Pos
    35  	bytesTmp         bytes.Buffer
    36  	matches          []hashAndMask // A hash matches if one of these matches.
    37  	excludes         []hashAndMask // explicitly excluded hash suffixes
    38  	bisect           *bisect.Matcher
    39  	fileSuffixOnly   bool // for Pos hashes, remove the directory prefix.
    40  	inlineSuffixOnly bool // for Pos hashes, remove all but the most inline position.
    41  }
    42  
    43  // SetInlineSuffixOnly controls whether hashing and reporting use the entire
    44  // inline position, or just the most-inline suffix.  Compiler debugging tends
    45  // to want the whole inlining, debugging user problems (loopvarhash, e.g.)
    46  // typically does not need to see the entire inline tree, there is just one
    47  // copy of the source code.
    48  func (d *HashDebug) SetInlineSuffixOnly(b bool) *HashDebug {
    49  	d.inlineSuffixOnly = b
    50  	return d
    51  }
    52  
    53  // The default compiler-debugging HashDebug, for "-d=gossahash=..."
    54  var hashDebug *HashDebug
    55  
    56  var FmaHash *HashDebug     // for debugging fused-multiply-add floating point changes
    57  var LoopVarHash *HashDebug // for debugging shared/private loop variable changes
    58  var PGOHash *HashDebug     // for debugging PGO optimization decisions
    59  
    60  // DebugHashMatchPkgFunc reports whether debug variable Gossahash
    61  //
    62  //  1. is empty (returns true; this is a special more-quickly implemented case of 4 below)
    63  //
    64  //  2. is "y" or "Y" (returns true)
    65  //
    66  //  3. is "n" or "N" (returns false)
    67  //
    68  //  4. does not explicitly exclude the sha1 hash of pkgAndName (see step 6)
    69  //
    70  //  5. is a suffix of the sha1 hash of pkgAndName (returns true)
    71  //
    72  //  6. OR
    73  //     if the (non-empty) value is in the regular language
    74  //     "(-[01]+/)+?([01]+(/[01]+)+?"
    75  //     (exclude..)(....include...)
    76  //     test the [01]+ exclude substrings, if any suffix-match, return false (4 above)
    77  //     test the [01]+ include substrings, if any suffix-match, return true
    78  //     The include substrings AFTER the first slash are numbered 0,1, etc and
    79  //     are named fmt.Sprintf("%s%d", varname, number)
    80  //     As an extra-special case for multiple failure search,
    81  //     an excludes-only string ending in a slash (terminated, not separated)
    82  //     implicitly specifies the include string "0/1", that is, match everything.
    83  //     (Exclude strings are used for automated search for multiple failures.)
    84  //     Clause 6 is not really intended for human use and only
    85  //     matters for failures that require multiple triggers.
    86  //
    87  // Otherwise it returns false.
    88  //
    89  // Unless Flags.Gossahash is empty, when DebugHashMatchPkgFunc returns true the message
    90  //
    91  //	"%s triggered %s\n", varname, pkgAndName
    92  //
    93  // is printed on the file named in environment variable GSHS_LOGFILE,
    94  // or standard out if that is empty.  "Varname" is either the name of
    95  // the variable or the name of the substring, depending on which matched.
    96  //
    97  // Typical use:
    98  //
    99  //  1. you make a change to the compiler, say, adding a new phase
   100  //
   101  //  2. it is broken in some mystifying way, for example, make.bash builds a broken
   102  //     compiler that almost works, but crashes compiling a test in run.bash.
   103  //
   104  //  3. add this guard to the code, which by default leaves it broken, but does not
   105  //     run the broken new code if Flags.Gossahash is non-empty and non-matching:
   106  //
   107  //     if !base.DebugHashMatch(ir.PkgFuncName(fn)) {
   108  //     return nil // early exit, do nothing
   109  //     }
   110  //
   111  //  4. rebuild w/o the bad code,
   112  //     GOCOMPILEDEBUG=gossahash=n ./all.bash
   113  //     to verify that you put the guard in the right place with the right sense of the test.
   114  //
   115  //  5. use github.com/dr2chase/gossahash to search for the error:
   116  //
   117  //     go install github.com/dr2chase/gossahash@latest
   118  //
   119  //     gossahash -- <the thing that fails>
   120  //
   121  //     for example: GOMAXPROCS=1 gossahash -- ./all.bash
   122  //
   123  //  6. gossahash should return a single function whose miscompilation
   124  //     causes the problem, and you can focus on that.
   125  func DebugHashMatchPkgFunc(pkg, fn string) bool {
   126  	return hashDebug.MatchPkgFunc(pkg, fn, nil)
   127  }
   128  
   129  func DebugHashMatchPos(pos src.XPos) bool {
   130  	return hashDebug.MatchPos(pos, nil)
   131  }
   132  
   133  // HasDebugHash returns true if Flags.Gossahash is non-empty, which
   134  // results in hashDebug being not-nil.  I.e., if !HasDebugHash(),
   135  // there is no need to create the string for hashing and testing.
   136  func HasDebugHash() bool {
   137  	return hashDebug != nil
   138  }
   139  
   140  // TODO: Delete when we switch to bisect-only.
   141  func toHashAndMask(s, varname string) hashAndMask {
   142  	l := len(s)
   143  	if l > 64 {
   144  		s = s[l-64:]
   145  		l = 64
   146  	}
   147  	m := ^(^uint64(0) << l)
   148  	h, err := strconv.ParseUint(s, 2, 64)
   149  	if err != nil {
   150  		Fatalf("Could not parse %s (=%s) as a binary number", varname, s)
   151  	}
   152  
   153  	return hashAndMask{name: varname, hash: h, mask: m}
   154  }
   155  
   156  // NewHashDebug returns a new hash-debug tester for the
   157  // environment variable ev.  If ev is not set, it returns
   158  // nil, allowing a lightweight check for normal-case behavior.
   159  func NewHashDebug(ev, s string, file io.Writer) *HashDebug {
   160  	if s == "" {
   161  		return nil
   162  	}
   163  
   164  	hd := &HashDebug{name: ev, logfile: file}
   165  	if !strings.Contains(s, "/") {
   166  		m, err := bisect.New(s)
   167  		if err != nil {
   168  			Fatalf("%s: %v", ev, err)
   169  		}
   170  		hd.bisect = m
   171  		return hd
   172  	}
   173  
   174  	// TODO: Delete remainder of function when we switch to bisect-only.
   175  	ss := strings.Split(s, "/")
   176  	// first remove any leading exclusions; these are preceded with "-"
   177  	i := 0
   178  	for len(ss) > 0 {
   179  		s := ss[0]
   180  		if len(s) == 0 || len(s) > 0 && s[0] != '-' {
   181  			break
   182  		}
   183  		ss = ss[1:]
   184  		hd.excludes = append(hd.excludes, toHashAndMask(s[1:], fmt.Sprintf("%s%d", "HASH_EXCLUDE", i)))
   185  		i++
   186  	}
   187  	// hash searches may use additional EVs with 0, 1, 2, ... suffixes.
   188  	i = 0
   189  	for _, s := range ss {
   190  		if s == "" {
   191  			if i != 0 || len(ss) > 1 && ss[1] != "" || len(ss) > 2 {
   192  				Fatalf("Empty hash match string for %s should be first (and only) one", ev)
   193  			}
   194  			// Special case of should match everything.
   195  			hd.matches = append(hd.matches, toHashAndMask("0", fmt.Sprintf("%s0", ev)))
   196  			hd.matches = append(hd.matches, toHashAndMask("1", fmt.Sprintf("%s1", ev)))
   197  			break
   198  		}
   199  		if i == 0 {
   200  			hd.matches = append(hd.matches, toHashAndMask(s, fmt.Sprintf("%s", ev)))
   201  		} else {
   202  			hd.matches = append(hd.matches, toHashAndMask(s, fmt.Sprintf("%s%d", ev, i-1)))
   203  		}
   204  		i++
   205  	}
   206  	return hd
   207  }
   208  
   209  // TODO: Delete when we switch to bisect-only.
   210  func (d *HashDebug) excluded(hash uint64) bool {
   211  	for _, m := range d.excludes {
   212  		if (m.hash^hash)&m.mask == 0 {
   213  			return true
   214  		}
   215  	}
   216  	return false
   217  }
   218  
   219  // TODO: Delete when we switch to bisect-only.
   220  func hashString(hash uint64) string {
   221  	hstr := ""
   222  	if hash == 0 {
   223  		hstr = "0"
   224  	} else {
   225  		for ; hash != 0; hash = hash >> 1 {
   226  			hstr = string('0'+byte(hash&1)) + hstr
   227  		}
   228  	}
   229  	if len(hstr) > 24 {
   230  		hstr = hstr[len(hstr)-24:]
   231  	}
   232  	return hstr
   233  }
   234  
   235  // TODO: Delete when we switch to bisect-only.
   236  func (d *HashDebug) match(hash uint64) *hashAndMask {
   237  	for i, m := range d.matches {
   238  		if (m.hash^hash)&m.mask == 0 {
   239  			return &d.matches[i]
   240  		}
   241  	}
   242  	return nil
   243  }
   244  
   245  // MatchPkgFunc returns true if either the variable used to create d is
   246  // unset, or if its value is y, or if it is a suffix of the base-two
   247  // representation of the hash of pkg and fn.  If the variable is not nil,
   248  // then a true result is accompanied by stylized output to d.logfile, which
   249  // is used for automated bug search.
   250  func (d *HashDebug) MatchPkgFunc(pkg, fn string, note func() string) bool {
   251  	if d == nil {
   252  		return true
   253  	}
   254  	// Written this way to make inlining likely.
   255  	return d.matchPkgFunc(pkg, fn, note)
   256  }
   257  
   258  func (d *HashDebug) matchPkgFunc(pkg, fn string, note func() string) bool {
   259  	hash := bisect.Hash(pkg, fn)
   260  	return d.matchAndLog(hash, func() string { return pkg + "." + fn }, note)
   261  }
   262  
   263  // MatchPos is similar to MatchPkgFunc, but for hash computation
   264  // it uses the source position including all inlining information instead of
   265  // package name and path.
   266  // Note that the default answer for no environment variable (d == nil)
   267  // is "yes", do the thing.
   268  func (d *HashDebug) MatchPos(pos src.XPos, desc func() string) bool {
   269  	if d == nil {
   270  		return true
   271  	}
   272  	// Written this way to make inlining likely.
   273  	return d.matchPos(Ctxt, pos, desc)
   274  }
   275  
   276  func (d *HashDebug) matchPos(ctxt *obj.Link, pos src.XPos, note func() string) bool {
   277  	return d.matchPosWithInfo(ctxt, pos, nil, note)
   278  }
   279  
   280  func (d *HashDebug) matchPosWithInfo(ctxt *obj.Link, pos src.XPos, info any, note func() string) bool {
   281  	hash := d.hashPos(ctxt, pos)
   282  	if info != nil {
   283  		hash = bisect.Hash(hash, info)
   284  	}
   285  	return d.matchAndLog(hash,
   286  		func() string {
   287  			r := d.fmtPos(ctxt, pos)
   288  			if info != nil {
   289  				r += fmt.Sprintf(" (%v)", info)
   290  			}
   291  			return r
   292  		},
   293  		note)
   294  }
   295  
   296  // MatchPosWithInfo is similar to MatchPos, but with additional information
   297  // that is included for hash computation, so it can distinguish multiple
   298  // matches on the same source location.
   299  // Note that the default answer for no environment variable (d == nil)
   300  // is "yes", do the thing.
   301  func (d *HashDebug) MatchPosWithInfo(pos src.XPos, info any, desc func() string) bool {
   302  	if d == nil {
   303  		return true
   304  	}
   305  	// Written this way to make inlining likely.
   306  	return d.matchPosWithInfo(Ctxt, pos, info, desc)
   307  }
   308  
   309  // matchAndLog is the core matcher. It reports whether the hash matches the pattern.
   310  // If a report needs to be printed, match prints that report to the log file.
   311  // The text func must be non-nil and should return a user-readable
   312  // representation of what was hashed. The note func may be nil; if non-nil,
   313  // it should return additional information to display to the user when this
   314  // change is selected.
   315  func (d *HashDebug) matchAndLog(hash uint64, text, note func() string) bool {
   316  	if d.bisect != nil {
   317  		enabled := d.bisect.ShouldEnable(hash)
   318  		if d.bisect.ShouldPrint(hash) {
   319  			disabled := ""
   320  			if !enabled {
   321  				disabled = " [DISABLED]"
   322  			}
   323  			var t string
   324  			if !d.bisect.MarkerOnly() {
   325  				t = text()
   326  				if note != nil {
   327  					if n := note(); n != "" {
   328  						t += ": " + n + disabled
   329  						disabled = ""
   330  					}
   331  				}
   332  			}
   333  			d.log(d.name, hash, strings.TrimSpace(t+disabled))
   334  		}
   335  		return enabled
   336  	}
   337  
   338  	// TODO: Delete rest of function body when we switch to bisect-only.
   339  	if d.excluded(hash) {
   340  		return false
   341  	}
   342  	if m := d.match(hash); m != nil {
   343  		d.log(m.name, hash, text())
   344  		return true
   345  	}
   346  	return false
   347  }
   348  
   349  // short returns the form of file name to use for d.
   350  // The default is the full path, but fileSuffixOnly selects
   351  // just the final path element.
   352  func (d *HashDebug) short(name string) string {
   353  	if d.fileSuffixOnly {
   354  		return filepath.Base(name)
   355  	}
   356  	return name
   357  }
   358  
   359  // hashPos returns a hash of the position pos, including its entire inline stack.
   360  // If d.inlineSuffixOnly is true, hashPos only considers the innermost (leaf) position on the inline stack.
   361  func (d *HashDebug) hashPos(ctxt *obj.Link, pos src.XPos) uint64 {
   362  	if d.inlineSuffixOnly {
   363  		p := ctxt.InnermostPos(pos)
   364  		return bisect.Hash(d.short(p.Filename()), p.Line(), p.Col())
   365  	}
   366  	h := bisect.Hash()
   367  	ctxt.AllPos(pos, func(p src.Pos) {
   368  		h = bisect.Hash(h, d.short(p.Filename()), p.Line(), p.Col())
   369  	})
   370  	return h
   371  }
   372  
   373  // fmtPos returns a textual formatting of the position pos, including its entire inline stack.
   374  // If d.inlineSuffixOnly is true, fmtPos only considers the innermost (leaf) position on the inline stack.
   375  func (d *HashDebug) fmtPos(ctxt *obj.Link, pos src.XPos) string {
   376  	format := func(p src.Pos) string {
   377  		return fmt.Sprintf("%s:%d:%d", d.short(p.Filename()), p.Line(), p.Col())
   378  	}
   379  	if d.inlineSuffixOnly {
   380  		return format(ctxt.InnermostPos(pos))
   381  	}
   382  	var stk []string
   383  	ctxt.AllPos(pos, func(p src.Pos) {
   384  		stk = append(stk, format(p))
   385  	})
   386  	return strings.Join(stk, "; ")
   387  }
   388  
   389  // log prints a match with the given hash and textual formatting.
   390  // TODO: Delete varname parameter when we switch to bisect-only.
   391  func (d *HashDebug) log(varname string, hash uint64, text string) {
   392  	d.mu.Lock()
   393  	defer d.mu.Unlock()
   394  
   395  	file := d.logfile
   396  	if file == nil {
   397  		if tmpfile := os.Getenv("GSHS_LOGFILE"); tmpfile != "" {
   398  			var err error
   399  			file, err = os.OpenFile(tmpfile, os.O_RDWR|os.O_CREATE|os.O_APPEND, 0666)
   400  			if err != nil {
   401  				Fatalf("could not open hash-testing logfile %s", tmpfile)
   402  				return
   403  			}
   404  		}
   405  		if file == nil {
   406  			file = os.Stdout
   407  		}
   408  		d.logfile = file
   409  	}
   410  
   411  	// Bisect output.
   412  	fmt.Fprintf(file, "%s %s\n", text, bisect.Marker(hash))
   413  
   414  	// Gossahash output.
   415  	// TODO: Delete rest of function when we switch to bisect-only.
   416  	fmt.Fprintf(file, "%s triggered %s %s\n", varname, text, hashString(hash))
   417  }
   418
View as plain text