Involved Source Files Package bisect can be used by compilers and other programs
to serve as a target for the bisect debugging tool.
See [golang.org/x/tools/cmd/bisect] for details about using the tool.
To be a bisect target, allowing bisect to help determine which of a set of independent
changes provokes a failure, a program needs to:
1. Define a way to accept a change pattern on its command line or in its environment.
The most common mechanism is a command-line flag.
The pattern can be passed to [New] to create a [Matcher], the compiled form of a pattern.
2. Assign each change a unique ID. One possibility is to use a sequence number,
but the most common mechanism is to hash some kind of identifying information
like the file and line number where the change might be applied.
[Hash] hashes its arguments to compute an ID.
3. Enable each change that the pattern says should be enabled.
The [Matcher.ShouldEnable] method answers this question for a given change ID.
4. Print a report identifying each change that the pattern says should be printed.
The [Matcher.ShouldPrint] method answers this question for a given change ID.
The report consists of one more lines on standard error or standard output
that contain a “match marker”. [Marker] returns the match marker for a given ID.
When bisect reports a change as causing the failure, it identifies the change
by printing the report lines with the match marker removed.
# Example Usage
A program starts by defining how it receives the pattern. In this example, we will assume a flag.
The next step is to compile the pattern:
m, err := bisect.New(patternFlag)
if err != nil {
log.Fatal(err)
}
Then, each time a potential change is considered, the program computes
a change ID by hashing identifying information (source file and line, in this case)
and then calls m.ShouldPrint and m.ShouldEnable to decide whether to
print and enable the change, respectively. The two can return different values
depending on whether bisect is trying to find a minimal set of changes to
disable or to enable to provoke the failure.
It is usually helpful to write a helper function that accepts the identifying information
and then takes care of hashing, printing, and reporting whether the identified change
should be enabled. For example, a helper for changes identified by a file and line number
would be:
func ShouldEnable(file string, line int) {
h := bisect.Hash(file, line)
if m.ShouldPrint(h) {
fmt.Fprintf(os.Stderr, "%v %s:%d\n", bisect.Marker(h), file, line)
}
return m.ShouldEnable(h)
}
Finally, note that New returns a nil Matcher when there is no pattern,
meaning that the target is not running under bisect at all,
so all changes should be enabled and none should be printed.
In that common case, the computation of the hash can be avoided entirely
by checking for m == nil first:
func ShouldEnable(file string, line int) bool {
if m == nil {
return false
}
h := bisect.Hash(file, line)
if m.ShouldPrint(h) {
fmt.Fprintf(os.Stderr, "%v %s:%d\n", bisect.Marker(h), file, line)
}
return m.ShouldEnable(h)
}
When the identifying information is expensive to format, this code can call
[Matcher.MarkerOnly] to find out whether short report lines containing only the
marker are permitted for a given run. (Bisect permits such lines when it is
still exploring the space of possible changes and will not be showing the
output to the user.) If so, the client can choose to print only the marker:
func ShouldEnable(file string, line int) bool {
if m == nil {
return false
}
h := bisect.Hash(file, line)
if m.ShouldPrint(h) {
if m.MarkerOnly() {
bisect.PrintMarker(os.Stderr)
} else {
fmt.Fprintf(os.Stderr, "%v %s:%d\n", bisect.Marker(h), file, line)
}
}
return m.ShouldEnable(h)
}
This specific helper – deciding whether to enable a change identified by
file and line number and printing about the change when necessary – is
provided by the [Matcher.FileLine] method.
Another common usage is deciding whether to make a change in a function
based on the caller's stack, to identify the specific calling contexts that the
change breaks. The [Matcher.Stack] method takes care of obtaining the stack,
printing it when necessary, and reporting whether to enable the change
based on that stack.
# Pattern Syntax
Patterns are generated by the bisect tool and interpreted by [New].
Users should not have to understand the patterns except when
debugging a target's bisect support or debugging the bisect tool itself.
The pattern syntax selecting a change is a sequence of bit strings
separated by + and - operators. Each bit string denotes the set of
changes with IDs ending in those bits, + is set addition, - is set subtraction,
and the expression is evaluated in the usual left-to-right order.
The special binary number “y” denotes the set of all changes,
standing in for the empty bit string.
In the expression, all the + operators must appear before all the - operators.
A leading + adds to an empty set. A leading - subtracts from the set of all
possible suffixes.
For example:
- “01+10” and “+01+10” both denote the set of changes
with IDs ending with the bits 01 or 10.
- “01+10-1001” denotes the set of changes with IDs
ending with the bits 01 or 10, but excluding those ending in 1001.
- “-01-1000” and “y-01-1000 both denote the set of all changes
with IDs not ending in 01 nor 1000.
- “0+1-01+001” is not a valid pattern, because all the + operators do not
appear before all the - operators.
In the syntaxes described so far, the pattern specifies the changes to
enable and report. If a pattern is prefixed by a “!”, the meaning
changes: the pattern specifies the changes to DISABLE and report. This
mode of operation is needed when a program passes with all changes
enabled but fails with no changes enabled. In this case, bisect
searches for minimal sets of changes to disable.
Put another way, the leading “!” inverts the result from [Matcher.ShouldEnable]
but does not invert the result from [Matcher.ShouldPrint].
As a convenience for manual debugging, “n” is an alias for “!y”,
meaning to disable and report all changes.
Finally, a leading “v” in the pattern indicates that the reports will be shown
to the user of bisect to describe the changes involved in a failure.
At the API level, the leading “v” causes [Matcher.Visible] to return true.
See the next section for details.
# Match Reports
The target program must enable only those changed matched
by the pattern, and it must print a match report for each such change.
A match report consists of one or more lines of text that will be
printed by the bisect tool to describe a change implicated in causing
a failure. Each line in the report for a given change must contain a
match marker with that change ID, as returned by [Marker].
The markers are elided when displaying the lines to the user.
A match marker has the form “[bisect-match 0x1234]” where
0x1234 is the change ID in hexadecimal.
An alternate form is “[bisect-match 010101]”, giving the change ID in binary.
When [Matcher.Visible] returns false, the match reports are only
being processed by bisect to learn the set of enabled changes,
not shown to the user, meaning that each report can be a match
marker on a line by itself, eliding the usual textual description.
When the textual description is expensive to compute,
checking [Matcher.Visible] can help the avoid that expense
in most runs.
Package-Level Type Names (total 6, in which 2 are exported)
/* sort exporteds by: | */
A Matcher is the parsed, compiled form of a PATTERN string.
The nil *Matcher is valid: it has all changes enabled but none reported.dedupatomicPointerDedup // when true, list is for “enable and report” (when false, “disable and report”) // conditions; later ones win over earlier ones // disables all reporting. reset if verbose is true. use case is -d=fmahash=qn // annotate reporting with human-helpful information FileLine reports whether the change identified by file and line should be enabled.
If the change should be printed, FileLine prints a one-line report to w. MarkerOnly reports whether it is okay to print only the marker for
a given change, omitting the identifying information.
MarkerOnly returns true when bisect is using the printed reports
only for an intermediate search step, not for showing to users. ShouldEnable reports whether the change with the given id should be enabled. ShouldPrint reports whether to print identifying information about the change with the given id. MatchStack assigns the current call stack a change ID.
If the stack should be printed, MatchStack prints it.
Then MatchStack reports whether a change at the current call stack should be enabled. fileLine does the real work for FileLine.
This lets FileLine's body handle m == nil and potentially be inlined. matchResult returns the result from the first condition that matches id. stack does the real work for Stack.
This lets stack's body handle m == nil and potentially be inlined.
func New(pattern string) (*Matcher, error)
atomicPointerDedup is an atomic.Pointer[dedup],
but we are avoiding using Go 1.19's atomic.Pointer
until the bootstrap toolchain can be relied upon to have it.punsafe.Pointer(*atomicPointerDedup) CompareAndSwap(old, new *dedup) bool(*atomicPointerDedup) Load() *dedup
A cond is a single condition in the matcher.
Given an input id, if id&mask == bits, return the result.bitsuint64maskuint64resultbool
A dedup is a deduplicator for call stacks, so that we only print
a report for new call stacks, not for call stacks we've already
reported.
It has two modes: an approximate but lock-free mode that
may still emit some duplicates, and a precise mode that uses
a lock and never emits duplicates.mmap[uint64]bool complete history for seen 128-entry 4-way, lossy cache for seenLossy seen records that h has now been seen and reports whether it was seen before.
When seen returns false, the caller is expected to print a report for h. seenLossy is a variant of seen that avoids a lock by using a cache of recently seen hashes.
Each cache entry is N-way set-associative: h can appear in any of the slots.
If h does not appear in any of them, then it is inserted into a random slot,
overwriting whatever was there before.
parseError is a trivial error implementation,
defined here to avoid importing errors.textstring(*parseError) Error() string
*parseError : error
Package-Level Functions (total 13, in which 6 are exported)
AppendMarker is like [Marker] but appends the marker to dst.
CutMarker finds the first match marker in line and removes it,
returning the shortened line (with the marker removed),
the ID from the match marker,
and whether a marker was found at all.
If there is no marker, CutMarker returns line, 0, false.
Hash computes a hash of the data arguments,
each of which must be of type string, byte, int, uint, int32, uint32, int64, uint64, uintptr, or a slice of one of those types.
Marker returns the match marker text to use on any line reporting details
about a match of the given ID.
It always returns the hexadecimal format.
New creates and returns a new Matcher implementing the given pattern.
The pattern syntax is defined in the package doc comment.
In addition to the pattern syntax syntax, New("") returns nil, nil.
The nil *Matcher is valid for use: it returns true from ShouldEnable
and false from ShouldPrint for all changes. Callers can avoid calling
[Hash], [Matcher.ShouldEnable], and [Matcher.ShouldPrint] entirely
when they recognize the nil Matcher.
PrintMarker prints to w a one-line report containing only the marker for h.
It is appropriate to use when [Matcher.ShouldPrint] and [Matcher.MarkerOnly] both return true.
appendFileLine appends file:line to dst, returning the extended slice.
The pages are generated with Goldsv0.6.7. (GOOS=linux GOARCH=amd64)
Golds is a Go 101 project developed by Tapir Liu.
PR and bug reports are welcome and can be submitted to the issue list.
Please follow @Go100and1 (reachable from the left QR code) to get the latest news of Golds.
