Allow "." to be excluded from @INC

[perl5.git] / Porting / bench.pl

#!/usr/bin/perl
#
# A tool for analysing the performance of the code snippets found in
# t/perf/benchmarks or similar


=head1 NAME

bench.pl - Compare the performance of perl code snippets across multiple
perls.

=head1 SYNOPSIS

    # Basic: run the tests in t/perf/benchmarks against two or
    # more perls

    bench.pl [options] -- perlA[=labelA] perlB[=labelB] ...

    # run the tests against same perlA 2x, with and without extra
    # options

    bench.pl [options] -- perlA=fast PerlA=slow -Mstrict -Dpsltoc 

    # Run bench.pl's own built-in sanity tests

    bench.pl --action=selftest

=head1 DESCRIPTION

By default, F<bench.pl> will run code snippets found in
F<t/perf/benchmarks> (or similar) under cachegrind, in order to calculate
how many instruction reads, data writes, branches, cache misses, etc. that
one execution of the snippet uses. It will run them against two or more
perl executables and show how much each test has gotten better or worse.

It is modelled on the F<perlbench> tool, but since it measures instruction
reads etc., rather than timings, it is much more precise and reproducible.
It is also considerably faster, and is capable of running tests in
parallel (with C<-j>). Rather than  displaying a single relative
percentage per test/perl combination, it displays values for 13 different
measurements, such as instruction reads, conditional branch misses etc.

There are options to write the raw data to a file, and to read it back.
This means that you can view the same run data in different views with
different selection and sort options.

The optional C<=label> after each perl executable is used in the display
output.

=head1 OPTIONS

=over 4

=item *

--action=I<foo>

What action to perform. The default is  I<grind>, which runs the benchmarks
using I<cachegrind> as the back end. The only other action at the moment is
I<selftest>, which runs some basic sanity checks and produces TAP output.

=item *

--average

Only display the overall average, rather than the results for each
individual test.

=item *

--benchfile=I<foo>

The path of the file which contains the benchmarks (F<t/perf/benchmarks>
by default).

=item *

--bisect=I<field,minval,maxval>

Run a single test against one perl and exit with a zero status if the
named field is in the specified range; exit 1 otherwise. It will complain
if more than one test or perl has been specified. It is intended to be
called as part of a bisect run, to determine when something changed.
For example,

    bench.pl -j 8 --tests=foo --bisect=Ir,100,105 --perlargs=-Ilib \
        ./miniperl

might be called from bisect to find when the number of instruction reads
for test I<foo> falls outside the range 100..105.

=item *

--compact=<Iperl>

Display the results for a single perl executable in a compact form.
Which perl to display is specified in the same manner as C<--norm>.

=item *

--debug

Enable verbose debugging output.

=item *

--fields=I<a,b,c>

Display only the specified fields; for example,

    --fields=Ir,Ir_m,Ir_mm

If only one field is selected, the output is in more compact form.

=item *

--grindargs=I<foo>

Optional command-line arguments to pass to all cachegrind invocations.

This option is appended to those which bench.pl uses for its own
purposes; so it can be used to override them (see --debug output
below), and can also be 'abused' to add redirects into the valgrind
command invocation.

For example, this writes PERL_MEM_LOG activity to foobar.$$, because
3>foobar.$$ redirects fd 3, then perl under PERL_MEM_LOG writes to fd 3.

 $ perl Porting/bench.pl --jobs=2 --verbose --debug \
    --tests=call::sub::amp_empty \
    \
    --grindargs='--cachegrind-out-file=junk.$$ 3>foobar.$$' \
    -- \
    perl5.24.0	perl5.24.0:+memlog:PERL_MEM_LOG=3mst

for the +memlog tests, this executes as: (shown via --debug, then prettyfied)

  Command: PERL_HASH_SEED=0 PERL_MEM_LOG=3mst
    valgrind --tool=cachegrind  --branch-sim=yes
    --cachegrind-out-file=/dev/null --cachegrind-out-file=junk.$$
    3>foobar.$$ perl5.24.0  - 10 2>&1

The result is that a set of junk.$$ files containing raw cachegrind
output are written, and foobar.$$ contains the expected memlog output.

Notes:

Theres no obvious utility for those junk.$$ and foobar.$$ files, but
you can have them anyway.

The 3 in PERL_MEM_LOG=3mst is needed because the output would
otherwize go to STDERR, and cause parse_cachegrind() to reject the
test and die.

The --grindargs redirect is needed to capture the memlog output;
without it, the memlog output is written to fd3, around
parse_cachegrind and effectively into /dev/null

PERL_MEM_LOG is expensive when used.

call::sub::amp_empty
&foo function call with no args or body

       perl5.24.0 perl5.24.0+memlog
       ---------- -----------------
    Ir      394.0          543477.5
    Dr      161.0          146814.1
    Dw       72.0          122304.6
  COND       58.0           66796.4
   IND        5.0            5537.7

COND_m        0.0            6743.1
 IND_m        5.0            1490.2

 Ir_m1        0.0             683.7
 Dr_m1        0.0              65.9
 Dw_m1        0.0               8.5

 Ir_mm        0.0              11.6
 Dr_mm        0.0              10.6
 Dw_mm        0.0               4.7


=item *

---help

Display basic usage information.

=item *

-j I<N>
--jobs=I<N>

Run I<N> jobs in parallel (default 1). This determines how many cachegrind
process will running at a time, and should generally be set to the number
of CPUs available.

=item *

--norm=I<foo>

Specify which perl column in the output to treat as the 100% norm.
It may be a column number (0..N-1) or a perl executable name or label.
It defaults to the leftmost column.

=item *

--perlargs=I<foo>

Optional command-line arguments to pass to each perl-under-test
(perlA, perlB in synopsis) For example, C<--perlargs=-Ilib>.

=item *

--raw

Display raw data counts rather than percentages in the outputs. This
allows you to see the exact number of intruction reads, branch misses etc.
for each test/perl combination. It also causes the C<AVERAGE> display
per field to be calculated based on the average of each tests's count
rather than average of each percentage. This means that tests with very
high counts will dominate.

=item *

--sort=I<field:perl>

Order the tests in the output based on the value of I<field> in the
column I<perl>. The I<perl> value is as per C<--norm>. For example

    bench.pl --sort=Dw:perl-5.20.0 \
        perl-5.16.0 perl-5.18.0 perl-5.20.0

=item *

-r I<file>
--read=I<file>

Read in saved data from a previous C<--write> run from the specified file.

Requires C<JSON::PP> to be available.

=item *

--tests=I<FOO>

Specify a subset of tests to run (or in the case of C<--read>, to display).
It may be either a comma-separated list of test names, or a regular
expression. For example

    --tests=expr::assign::scalar_lex,expr::assign::2list_lex
    --tests=/^expr::/

=item *

--verbose

Display progress information.

=item *

-w I<file>
--write=I<file>

Save the raw data to the specified file. It can be read back later with
C<--read>.

Requires C<JSON::PP> to be available.

=back

=cut



use 5.010000;
use warnings;
use strict;
use Getopt::Long qw(:config no_auto_abbrev require_order);
use IPC::Open2 ();
use IO::Select;
use IO::File;
use POSIX ":sys_wait_h";

# The version of the file format used to save data. We refuse to process
# the file if the integer component differs.

my $FORMAT_VERSION = 1.0;

# The fields we know about

my %VALID_FIELDS = map { $_ => 1 }
    qw(Ir Ir_m1 Ir_mm Dr Dr_m1 Dr_mm Dw Dw_m1 Dw_mm COND COND_m IND IND_m);

sub usage {
    die <<EOF;
usage: $0 [options] -- perl[=label] ...
  --action=foo       What action to perform [default: grind].
  --average          Only display average, not individual test results.
  --benchfile=foo    File containing the benchmarks;
                       [default: t/perf/benchmarks].
  --bisect=f,min,max run a single test against one perl and exit with a
                       zero status if the named field is in the specified
                       range; exit 1 otherwise.
  --compact=perl     Display the results of a single perl in compact form.
                     Which perl specified like --norm
  --debug            Enable verbose debugging output.
  --fields=a,b,c     Display only the specified fields (e.g. Ir,Ir_m,Ir_mm).
  --grindargs=foo    Optional command-line args to pass to cachegrind.
  --help             Display this help.
  -j|--jobs=N        Run N jobs in parallel [default 1].
  --norm=perl        Which perl column to treat as 100%; may be a column
                       number (0..N-1) or a perl executable name or label;
                       [default: 0].
  --perlargs=foo     Optional command-line args to pass to each perl to run.
  --raw              Display raw data counts rather than percentages.
  --sort=field:perl  Sort the tests based on the value of 'field' in the
                       column 'perl'. The perl value is as per --norm.
  -r|--read=file     Read in previously saved data from the specified file.
  --tests=FOO        Select only the specified tests from the benchmarks file;
                       FOO may be either of the form 'foo,bar' or '/regex/';
                       [default: all tests].
  --verbose          Display progress information.
  -w|--write=file    Save the raw data to the specified file.

--action is one of:
    grind            run the code under cachegrind
    selftest         perform a selftest; produce TAP output

The command line ends with one or more specified perl executables,
which will be searched for in the current \$PATH. Each binary name may
have an optional =LABEL appended, which will be used rather than the
executable name in output. E.g.

    perl-5.20.1=PRE-BUGFIX  perl-5.20.1-new=POST-BUGFIX
EOF
}

my %OPTS = (
    action    => 'grind',
    average   => 0,
    benchfile => 't/perf/benchmarks',
    bisect    => undef,
    compact   => undef,
    debug     => 0,
    grindargs => '',
    fields    => undef,
    jobs      => 1,
    norm      => 0,
    perlargs  => '',
    raw       => 0,
    read      => undef,
    sort      => undef,
    tests     => undef,
    verbose   => 0,
    write     => undef,
);


# process command-line args and call top-level action

{
    GetOptions(
        'action=s'    => \$OPTS{action},
        'average'     => \$OPTS{average},
        'benchfile=s' => \$OPTS{benchfile},
        'bisect=s'    => \$OPTS{bisect},
        'compact=s'   => \$OPTS{compact},
        'debug'       => \$OPTS{debug},
        'grindargs=s' => \$OPTS{grindargs},
        'help'        => \$OPTS{help},
        'fields=s'    => \$OPTS{fields},
        'jobs|j=i'    => \$OPTS{jobs},
        'norm=s'      => \$OPTS{norm},
        'perlargs=s'  => \$OPTS{perlargs},
        'raw'         => \$OPTS{raw},
        'read|r=s'    => \$OPTS{read},
        'sort=s'      => \$OPTS{sort},
        'tests=s'     => \$OPTS{tests},
        'verbose'     => \$OPTS{verbose},
        'write|w=s'   => \$OPTS{write},
    ) or usage;

    usage if $OPTS{help};


    if (defined $OPTS{read} and defined $OPTS{write}) {
        die "Error: can't specify both --read and --write options\n";
    }

    if (defined $OPTS{read} or defined $OPTS{write}) {
        # fail early if it's not present
        require JSON::PP;
    }

    if (defined $OPTS{fields}) {
        my @f = split /,/, $OPTS{fields};
        for (@f) {
            die "Error: --fields: unknown field '$_'\n"
                unless $VALID_FIELDS{$_};
        }
        my %f = map { $_ => 1 } @f;
        $OPTS{fields} = \%f;
    }

    my %valid_actions = qw(grind 1 selftest 1);
    unless ($valid_actions{$OPTS{action}}) {
        die "Error: unrecognised action '$OPTS{action}'\n"
          . "must be one of: " . join(', ', sort keys %valid_actions)."\n";
    }

    if (defined $OPTS{sort}) {
        my @s = split /:/, $OPTS{sort};
        if (@s != 2) {
            die "Error: --sort argument should be of the form field:perl: "
              . "'$OPTS{sort}'\n";
        }
        my ($field, $perl) = @s;
        die "Error: --sort: unknown field '$field\n"
            unless $VALID_FIELDS{$field};
        # the 'perl' value will be validated later, after we have processed
        # the perls
        $OPTS{'sort-field'} = $field;
        $OPTS{'sort-perl'}  = $perl;
    }

    if ($OPTS{action} eq 'selftest') {
        if (@ARGV) {
            die "Error: no perl executables may be specified with --read\n"
        }
    }
    elsif (defined $OPTS{bisect}) {
        die "Error: exactly one perl executable must be specified for bisect\n"
                                                unless @ARGV == 1;
        die "Error: Can't specify both --bisect and --read\n"
                                                if defined $OPTS{read};
        die "Error: Can't specify both --bisect and --write\n"
                                                if defined $OPTS{write};
    }
    elsif (defined $OPTS{read}) {
        if (@ARGV) {
            die "Error: no perl executables may be specified with --read\n"
        }
    }
    elsif ($OPTS{raw}) {
        unless (@ARGV) {
            die "Error: at least one perl executable must be specified\n";
        }
    }
    else {
        unless (@ARGV >= 2) {
            die "Error: at least two perl executables must be specified\n";
        }
    }

    if ($OPTS{action} eq 'grind') {
        do_grind(\@ARGV);
    }
    elsif ($OPTS{action} eq 'selftest') {
        do_selftest();
    }
}
exit 0;


# Given a hash ref keyed by test names, filter it by deleting unwanted
# tests, based on $OPTS{tests}.

sub filter_tests {
    my ($tests) = @_;

    my $opt = $OPTS{tests};
    return unless defined $opt;

    my @tests;

    if ($opt =~ m{^/}) {
        $opt =~ s{^/(.+)/$}{$1}
            or die "Error: --tests regex must be of the form /.../\n";
        for (keys %$tests) {
            delete $tests->{$_} unless /$opt/;
        }
    }
    else {
        my %t;
        for (split /,/, $opt) {
            die "Error: no such test found: '$_'\n"
                . ($OPTS{verbose} ? "  have: @{[ sort keys %$tests ]}\n" : "")
                unless exists $tests->{$_};
            $t{$_} = 1;
        }
        for (keys %$tests) {
            delete $tests->{$_} unless exists $t{$_};
        }
    }
}


# Read in the test file, and filter out any tests excluded by $OPTS{tests}
# return a hash ref { testname => { test }, ... }
# and an array ref of the original test names order,

sub read_tests_file {
    my ($file) = @_;

    my $ta = do $file;
    unless ($ta) {
        die "Error: can't parse '$file': $@\n" if $@;
        die "Error: can't read '$file': $!\n";
    }

    my @orig_order;
    for (my $i=0; $i < @$ta; $i += 2) {
        push @orig_order, $ta->[$i];
    }

    my $t = { @$ta };
    filter_tests($t);
    return $t, \@orig_order;
}


# Process the perl/column argument of options like --norm and --sort.
# Return the index of the matching perl.

sub select_a_perl {
    my ($perl, $perls, $who) = @_;

    if ($perl =~ /^[0-9]$/) {
        die "Error: $who value $perl outside range 0.." . $#$perls . "\n"
                                        unless $perl < @$perls;
        return $perl;
    }
    else {
        my @perl = grep    $perls->[$_][0] eq $perl
                        || $perls->[$_][1] eq $perl,
                        0..$#$perls;
        die "Error: $who: unrecognised perl '$perl'\n"
                                        unless @perl;
        die "Error: $who: ambiguous perl '$perl'\n"
                                        if @perl > 1;
        return $perl[0];
    }
}


# Validate the list of perl=label (+ cmdline options) on the command line.
# Return a list of [ exe, label, cmdline-options ] tuples, i.e.
# 'perl-under-test's (PUTs)

sub process_puts {
    my @res_puts; # returned, each item is [ perlexe, label, @putargs ]
    my %seen;
    my @putargs; # collect not-perls into args per PUT

    for my $p (reverse @_) {
        push @putargs, $p and next if $p =~ /^-/; # not-perl, dont send to qx//

        my ($perl, $label, $env) = split /[=:,]/, $p, 3;
        $label //= $perl;
        $label = $perl.$label if $label =~ /^\+/;
        die "$label cannot be used on 2 different perls under test\n" if $seen{$label}++;

        my %env;
        if ($env) {
            %env = split /[=,]/, $env;
        }
        my $r = qx($perl -e 'print qq(ok\n)' 2>&1);
        if ($r eq "ok\n") {
	    push @res_puts, [ $perl, $label, \%env, reverse @putargs ];
            @putargs = ();
            warn "Added Perl-Under-Test: [ @{[@{$res_puts[-1]}]} ]\n"
                if $OPTS{verbose};
	} else {
            warn "perl-under-test args: @putargs + a not-perl: $p $r\n"
                if $OPTS{verbose};
            push @putargs, $p; # not-perl
	}
    }
    return reverse @res_puts;
}



# Return a string containing perl test code wrapped in a loop
# that runs $ARGV[0] times

sub make_perl_prog {
    my ($test, $desc, $setup, $code) = @_;

    return <<EOF;
# $desc
package $test;
BEGIN { srand(0) }
$setup;
for my \$__loop__ (1..\$ARGV[0]) {
    $code;
}
EOF
}


# Parse the output from cachegrind. Return a hash ref.
# See do_selftest() for examples of the output format.

sub parse_cachegrind {
    my ($output, $id, $perl) = @_;

    my %res;

    my @lines = split /\n/, $output;
    for (@lines) {
        unless (s/(==\d+==)|(--\d+--) //) {
            die "Error: while executing $id:\n"
              . "unexpected code or cachegrind output:\n$_\n";
        }
        if (/I   refs:\s+([\d,]+)/) {
            $res{Ir} = $1;
        }
        elsif (/I1  misses:\s+([\d,]+)/) {
            $res{Ir_m1} = $1;
        }
        elsif (/LLi misses:\s+([\d,]+)/) {
            $res{Ir_mm} = $1;
        }
        elsif (/D   refs:\s+.*?([\d,]+) rd .*?([\d,]+) wr/) {
            @res{qw(Dr Dw)} = ($1,$2);
        }
        elsif (/D1  misses:\s+.*?([\d,]+) rd .*?([\d,]+) wr/) {
            @res{qw(Dr_m1 Dw_m1)} = ($1,$2);
        }
        elsif (/LLd misses:\s+.*?([\d,]+) rd .*?([\d,]+) wr/) {
            @res{qw(Dr_mm Dw_mm)} = ($1,$2);
        }
        elsif (/Branches:\s+.*?([\d,]+) cond .*?([\d,]+) ind/) {
            @res{qw(COND IND)} = ($1,$2);
        }
        elsif (/Mispredicts:\s+.*?([\d,]+) cond .*?([\d,]+) ind/) {
            @res{qw(COND_m IND_m)} = ($1,$2);
        }
    }

    for my $field (keys %VALID_FIELDS) {
        die "Error: can't parse '$field' field from cachegrind output:\n$output"
            unless exists $res{$field};
        $res{$field} =~ s/,//g;
    }

    return \%res;
}


# Handle the 'grind' action

sub do_grind {
    my ($perl_args) = @_; # the residue of @ARGV after option processing

    my ($loop_counts, $perls, $results, $tests, $order);
    my ($bisect_field, $bisect_min, $bisect_max);

    if (defined $OPTS{bisect}) {
        ($bisect_field, $bisect_min, $bisect_max) = split /,/, $OPTS{bisect}, 3;
        die "Error: --bisect option must be of form 'field,integer,integer'\n"
            unless
                    defined $bisect_max
                and $bisect_min =~ /^[0-9]+$/
                and $bisect_max =~ /^[0-9]+$/;

        die "Error: unrecognised field '$bisect_field' in --bisect option\n"
            unless $VALID_FIELDS{$bisect_field};

        die "Error: --bisect min ($bisect_min) must be <= max ($bisect_max)\n"
            if $bisect_min > $bisect_max;
    }

    if (defined $OPTS{read}) {
        open my $in, '<:encoding(UTF-8)', $OPTS{read}
            or die " Error: can't open $OPTS{read} for reading: $!\n";
        my $data = do { local $/; <$in> };
        close $in;

        my $hash = JSON::PP::decode_json($data);
        if (int($FORMAT_VERSION) < int($hash->{version})) {
            die "Error: unsupported version $hash->{version} in file"
              . "'$OPTS{read}' (too new)\n";
        }
        ($loop_counts, $perls, $results, $tests, $order) =
            @$hash{qw(loop_counts perls results tests order)};

        filter_tests($results);
        filter_tests($tests);

        if (!$order) {
            $order = [ sort keys %$tests ];
        }
    }
    else {
        # How many times to execute the loop for the two trials. The lower
        # value is intended to do the loop enough times that branch
        # prediction has taken hold; the higher loop allows us to see the
        # branch misses after that
        $loop_counts = [10, 20];

        ($tests, $order) = read_tests_file($OPTS{benchfile});
        die "Error: only a single test may be specified with --bisect\n"
            if defined $OPTS{bisect} and keys %$tests != 1;

        $perls = [ process_puts(@$perl_args) ];


        $results = grind_run($tests, $order, $perls, $loop_counts);
    }

    # now that we have a list of perls, use it to process the
    # 'perl' component of the --norm and --sort args

    $OPTS{norm} = select_a_perl($OPTS{norm}, $perls, "--norm");
    if (defined $OPTS{'sort-perl'}) {
        $OPTS{'sort-perl'} =
                select_a_perl($OPTS{'sort-perl'}, $perls, "--sort");
    }

    if (defined $OPTS{'compact'}) {
        $OPTS{'compact'} =
                select_a_perl($OPTS{'compact'}, $perls, "--compact");
    }
    if (defined $OPTS{write}) {
        my $json = JSON::PP::encode_json({
                    version      => $FORMAT_VERSION,
                    loop_counts  => $loop_counts,
                    perls        => $perls,
                    results      => $results,
                    tests        => $tests,
                    order        => $order,
                });

        open my $out, '>:encoding(UTF-8)', $OPTS{write}
            or die " Error: can't open $OPTS{write} for writing: $!\n";
        print $out $json or die "Error: writing to file '$OPTS{write}': $!\n";
        close $out       or die "Error: closing file '$OPTS{write}': $!\n";
    }
    else {
        my ($processed, $averages) =
                    grind_process($results, $perls, $loop_counts);

        if (defined $OPTS{bisect}) {
            my @r = values %$results;
            die "Panic: expected exactly one test result in bisect\n"
                                                            if @r != 1;
            @r = values %{$r[0]};
            die "Panic: expected exactly one perl result in bisect\n"
                                                            if @r != 1;
            my $c = $r[0]{$bisect_field};
            die "Panic: no result in bisect for field '$bisect_field'\n"
                                                            unless defined $c;
            exit 0 if $bisect_min <= $c and $c <= $bisect_max;
            exit 1;
        }
        elsif (defined $OPTS{compact}) {
            grind_print_compact($processed, $averages, $OPTS{compact},
                                $perls, $tests, $order);
        }
        else {
            grind_print($processed, $averages, $perls, $tests, $order);
        }
    }
}


# Run cachegrind for every test/perl combo.
# It may run several processes in parallel when -j is specified.
# Return a hash ref suitable for input to grind_process()

sub grind_run {
    my ($tests, $order, $perls, $counts) = @_;

    # Build a list of all the jobs to run

    my @jobs;

    for my $test (grep $tests->{$_}, @$order) {

        # Create two test progs: one with an empty loop and one with code.
        # Note that the empty loop is actually '{1;}' rather than '{}';
        # this causes the loop to have a single nextstate rather than a
        # stub op, so more closely matches the active loop; e.g.:
        #   {1;}    => nextstate;                       unstack
        #   {$x=1;} => nextstate; const; gvsv; sassign; unstack
        my @prog = (
            make_perl_prog($test, @{$tests->{$test}}{qw(desc setup)}, '1'),
            make_perl_prog($test, @{$tests->{$test}}{qw(desc setup code)}),
        );

        for my $p (@$perls) {
            my ($perl, $label, $env, @putargs) = @$p;

            # Run both the empty loop and the active loop
            # $counts->[0] and $counts->[1] times.

            for my $i (0,1) {
                for my $j (0,1) {
                    my $envstr = '';
                    if (ref $env) {
                        $envstr .= "$_=$env->{$_} " for sort keys %$env;
                    }
                    my $cmd = "PERL_HASH_SEED=0 $envstr"
                            . "valgrind --tool=cachegrind  --branch-sim=yes "
                            . "--cachegrind-out-file=/dev/null "
                            . "$OPTS{grindargs} "
                            . "$perl $OPTS{perlargs} @putargs - $counts->[$j] 2>&1";
                    # for debugging and error messages
                    my $id = "$test/$label "
                        . ($i ? "active" : "empty") . "/"
                        . ($j ? "long"   : "short") . " loop";

                    push @jobs, {
                        test   => $test,
                        perl   => $perl,
                        plabel => $label,
                        cmd    => $cmd,
                        prog   => $prog[$i],
                        active => $i,
                        loopix => $j,
                        id     => $id,
                    };
                }
            }
        }
    }

    # Execute each cachegrind and store the results in %results.

    local $SIG{PIPE} = 'IGNORE';

    my $max_jobs = $OPTS{jobs};
    my $running  = 0; # count of executing jobs
    my %pids;         # map pids to jobs
    my %fds;          # map fds  to jobs
    my %results;
    my $select = IO::Select->new();

    while (@jobs or $running) {

        if ($OPTS{debug}) {
            printf "Main loop: pending=%d running=%d\n",
                scalar(@jobs), $running;
        }

        # Start new jobs

        while (@jobs && $running < $max_jobs) {
            my $job = shift @jobs;
            my ($id, $cmd) =@$job{qw(id cmd)};

            my ($in, $out, $pid);
            warn "Starting $id\n" if $OPTS{verbose};
            eval { $pid = IPC::Open2::open2($out, $in, $cmd); 1; }
                or die "Error: while starting cachegrind subprocess"
                   ." for $id:\n$@";
            $running++;
            $pids{$pid}    = $job;
            $fds{"$out"}   = $job;
            $job->{out_fd} = $out;
            $job->{output} = '';
            $job->{pid}    = $pid;

            $out->blocking(0);
            $select->add($out);

            if ($OPTS{debug}) {
                print "Started pid $pid for $id\n";
            }

            # Note:
            # In principle we should write to $in in the main select loop,
            # since it may block. In reality,
            #  a) the code we write to the perl process's stdin is likely
            #     to be less than the OS's pipe buffer size;
            #  b) by the time the perl process has read in all its stdin,
            #     the only output it should have generated is a few lines
            #     of cachegrind output preamble.
            # If these assumptions change, then perform the following print
            # in the select loop instead.

            print $in $job->{prog};
            close $in;
        }

        # Get output of running jobs

        if ($OPTS{debug}) {
            printf "Select: waiting on (%s)\n",
                join ', ', sort { $a <=> $b } map $fds{$_}{pid},
                            $select->handles;
        }

        my @ready = $select->can_read;

        if ($OPTS{debug}) {
            printf "Select: pids (%s) ready\n",
                join ', ', sort { $a <=> $b } map $fds{$_}{pid}, @ready;
        }

        unless (@ready) {
            die "Panic: select returned no file handles\n";
        }

        for my $fd (@ready) {
            my $j = $fds{"$fd"};
            my $r = sysread $fd, $j->{output}, 8192, length($j->{output});
            unless (defined $r) {
                die "Panic: Read from process running $j->{id} gave:\n$!";
            }
            next if $r;

            # EOF

            if ($OPTS{debug}) {
                print "Got eof for pid $fds{$fd}{pid} ($j->{id})\n";
            }

            $select->remove($j->{out_fd});
            close($j->{out_fd})
                or die "Panic: closing output fh on $j->{id} gave:\n$!\n";
            $running--;
            delete $fds{"$j->{out_fd}"};
            my $output = $j->{output};

            if ($OPTS{debug}) {
                my $p = $j->{prog};
                $p =~ s/^/    : /mg;
                my $o = $output;
                $o =~ s/^/    : /mg;

                print "\n$j->{id}/\nCommand: $j->{cmd}\n"
                    . "Input:\n$p"
                    . "Output\n$o";
            }

            $results{$j->{test}}{$j->{plabel}}[$j->{active}][$j->{loopix}]
                    = parse_cachegrind($output, $j->{id}, $j->{perl});
        }

        # Reap finished jobs

        while (1) {
            my $kid = waitpid(-1, WNOHANG);
            my $ret = $?;
            last if $kid <= 0;

            unless (exists $pids{$kid}) {
                die "Panic: reaped unexpected child $kid";
            }
            my $j = $pids{$kid};
            if ($ret) {
                die sprintf("Error: $j->{id} gave return status 0x%04x\n", $ret)
                    . "with the following output\n:$j->{output}\n";
            }
            delete $pids{$kid};
        }
    }

    return \%results;
}




# grind_process(): process the data that has been extracted from
# cachgegrind's output.
#
# $res is of the form ->{benchmark_name}{perl_name}[active][count]{field_name},
# where active is 0 or 1 indicating an empty or active loop,
# count is 0 or 1 indicating a short or long loop. E.g.
#
#    $res->{'expr::assign::scalar_lex'}{perl-5.21.1}[0][10]{Dw_mm}
#
# The $res data structure is modified in-place by this sub.
#
# $perls is [ [ perl-exe, perl-label], .... ].
#
# $counts is [ N, M ] indicating the counts for the short and long loops.
#
#
# return \%output, \%averages, where
#
# $output{benchmark_name}{perl_name}{field_name} = N
# $averages{perl_name}{field_name} = M
#
# where N is the raw count ($OPTS{raw}), or count_perl0/count_perlI otherwise;
# M is the average raw count over all tests ($OPTS{raw}), or
# 1/(sum(count_perlI/count_perl0)/num_tests) otherwise.

sub grind_process {
    my ($res, $perls, $counts) = @_;

    # Process the four results for each test/perf combo:
    # Convert
    #    $res->{benchmark_name}{perl_name}[active][count]{field_name} = n
    # to
    #    $res->{benchmark_name}{perl_name}{field_name} = averaged_n
    #
    # $r[0][1] - $r[0][0] is the time to do ($counts->[1]-$counts->[0])
    #                     empty loops, eliminating startup time
    # $r[1][1] - $r[1][0] is the time to do ($counts->[1]-$counts->[0])
    #                     active loops, eliminating startup time
    # (the two startup times may be different because different code
    # is being compiled); the difference of the two results above
    # divided by the count difference is the time to execute the
    # active code once, eliminating both startup and loop overhead.

    for my $tests (values %$res) {
        for my $r (values %$tests) {
            my $r2;
            for (keys %{$r->[0][0]}) {
                my $n = (  ($r->[1][1]{$_} - $r->[1][0]{$_})
                         - ($r->[0][1]{$_} - $r->[0][0]{$_})
                        ) / ($counts->[1] - $counts->[0]);
                $r2->{$_} = $n;
            }
            $r = $r2;
        }
    }

    my %totals;
    my %counts;
    my %data;

    my $perl_norm = $perls->[$OPTS{norm}][1]; # the label of the reference perl

    for my $test_name (keys %$res) {
        my $res1 = $res->{$test_name};
        my $res2_norm = $res1->{$perl_norm};
        for my $perl (keys %$res1) {
            my $res2 = $res1->{$perl};
            for my $field (keys %$res2) {
                my ($p, $q) = ($res2_norm->{$field}, $res2->{$field});

                if ($OPTS{raw}) {
                    # Avoid annoying '-0.0' displays. Ideally this number
                    # should never be negative, but fluctuations in
                    # startup etc can theoretically make this happen
                    $q = 0 if ($q <= 0 && $q > -0.1);
                    $totals{$perl}{$field} += $q;
                    $counts{$perl}{$field}++;
                    $data{$test_name}{$perl}{$field} = $q;
                    next;
                }

                # $p and $q are notionally integer counts, but
                # due to variations in startup etc, it's possible for a
                # count which is supposedly zero to be calculated as a
                # small positive or negative value.
                # In this case, set it to zero. Further below we
                # special-case zeros to avoid division by zero errors etc.

                $p = 0.0 if $p < 0.01;
                $q = 0.0 if $q < 0.01;

                if ($p == 0.0 && $q == 0.0) {
                    # Both perls gave a count of zero, so no change:
                    # treat as 100%
                    $totals{$perl}{$field} += 1;
                    $counts{$perl}{$field}++;
                    $data{$test_name}{$perl}{$field} = 1;
                }
                elsif ($p == 0.0 || $q == 0.0) {
                    # If either count is zero, there were too few events
                    # to give a meaningful ratio (and we will end up with
                    # division by zero if we try). Mark the result undef,
                    # indicating that it shouldn't be displayed; and skip
                    # adding to the average
                    $data{$test_name}{$perl}{$field} = undef;
                }
                else {
                    # For averages, we record q/p rather than p/q.
                    # Consider a test where perl_norm took 1000 cycles
                    # and perlN took 800 cycles. For the individual
                    # results we display p/q, or 1.25; i.e. a quarter
                    # quicker. For the averages, we instead sum all
                    # the 0.8's, which gives the total cycles required to
                    # execute all tests, with all tests given equal
                    # weight. Later we reciprocate the final result,
                    # i.e. 1/(sum(qi/pi)/n)

                    $totals{$perl}{$field} += $q/$p;
                    $counts{$perl}{$field}++;
                    $data{$test_name}{$perl}{$field} = $p/$q;
                }
            }
        }
    }

    # Calculate averages based on %totals and %counts accumulated earlier.

    my %averages;
    for my $perl (keys %totals) {
        my $t = $totals{$perl};
        for my $field (keys %$t) {
            $averages{$perl}{$field} = $OPTS{raw}
                ? $t->{$field} / $counts{$perl}{$field}
                  # reciprocal - see comments above
                : $counts{$perl}{$field} / $t->{$field};
        }
    }

    return \%data, \%averages;
}



# print a standard blurb at the start of the grind display

sub grind_blurb {
    my ($perls) = @_;

    print <<EOF;
Key:
    Ir   Instruction read
    Dr   Data read
    Dw   Data write
    COND conditional branches
    IND  indirect branches
    _m   branch predict miss
    _m1  level 1 cache miss
    _mm  last cache (e.g. L3) miss
    -    indeterminate percentage (e.g. 1/0)

EOF

    if ($OPTS{raw}) {
        print "The numbers represent raw counts per loop iteration.\n";
    }
    else {
        print <<EOF;
The numbers represent relative counts per loop iteration, compared to
$perls->[$OPTS{norm}][1] at 100.0%.
Higher is better: for example, using half as many instructions gives 200%,
while using twice as many gives 50%.
EOF
    }
}


# return a sorted list of the test names, plus 'AVERAGE'

sub sorted_test_names {
    my ($results, $order, $perls) = @_;

    my @names;
    unless ($OPTS{average}) {
        if (defined $OPTS{'sort-field'}) {
            my ($field, $perlix) = @OPTS{'sort-field', 'sort-perl'};
            my $perl = $perls->[$perlix][1];
            @names = sort
                {
                        $results->{$a}{$perl}{$field}
                    <=> $results->{$b}{$perl}{$field}
                }
                keys %$results;
        }
        else {
            @names = grep $results->{$_}, @$order;
        }
    }

    # No point in displaying average for only one test.
    push @names,  'AVERAGE' unless @names == 1;
    @names;
}


# grind_print(): display the tabulated results of all the cachegrinds.
#
# Arguments are of the form:
#    $results->{benchmark_name}{perl_name}{field_name} = N
#    $averages->{perl_name}{field_name} = M
#    $perls = [ [ perl-exe, perl-label ], ... ]
#    $tests->{test_name}{desc => ..., ...}

sub grind_print {
    my ($results, $averages, $perls, $tests, $order) = @_;

    my @perl_names = map $_->[0], @$perls;
    my @perl_labels = map $_->[1], @$perls;
    my %perl_labels;
    $perl_labels{$_->[0]} = $_->[1] for @$perls;

    my $field_label_width = 6;
    # Calculate the width to display for each column.
    my $min_width = $OPTS{raw} ? 8 : 6;
    my @widths = map { length($_) < $min_width ? $min_width : length($_) }
    			@perl_labels;

    # Print standard header.
    grind_blurb($perls);

    my @test_names = sorted_test_names($results, $order, $perls);

    # If only a single field is to be displayed, use a more compact
    # format with only a single line of output per test.

    my $one_field = defined $OPTS{fields} &&  keys(%{$OPTS{fields}}) == 1;

    if ($one_field) {
        print "Results for field " . (keys(%{$OPTS{fields}}))[0] . ".\n";

        # The first column will now contain test names rather than
        # field names; Calculate the max width.

        $field_label_width = 0;
        for (@test_names) {
            $field_label_width = length if length > $field_label_width;
        }

        # Print the perl executables header.

        print "\n";
        for my $i (0,1) {
            print " " x $field_label_width;
            for (0..$#widths) {
                printf " %*s", $widths[$_],
                    $i ? ('-' x$widths[$_]) :  $perl_labels[$_];
            }
            print "\n";
        }
    }

    # Dump the results for each test.

    for my $test_name (@test_names) {
        my $doing_ave = ($test_name eq 'AVERAGE');
        my $res1 = $doing_ave ? $averages : $results->{$test_name};

        unless ($one_field) {
            print "\n$test_name";
            print "\n$tests->{$test_name}{desc}" unless $doing_ave;
            print "\n\n";

            # Print the perl executables header.
            for my $i (0,1) {
                print " " x $field_label_width;
                for (0..$#widths) {
                    printf " %*s", $widths[$_],
                        $i ? ('-' x$widths[$_]) :  $perl_labels[$_];
                }
                print "\n";
            }
        }

        for my $field (qw(Ir Dr Dw COND IND
                          N
                          COND_m IND_m
                          N
                          Ir_m1 Dr_m1 Dw_m1
                          N
                          Ir_mm Dr_mm Dw_mm
                      ))
        {
            next if $OPTS{fields} and ! exists $OPTS{fields}{$field};

            if ($field eq 'N') {
                print "\n";
                next;
            }

            if ($one_field) {
                printf "%-*s", $field_label_width, $test_name;
            }
            else {
                printf "%*s", $field_label_width, $field;
            }

            for my $i (0..$#widths) {
                my $res2 = $res1->{$perl_labels[$i]};
                my $p = $res2->{$field};
                if (!defined $p) {
                    printf " %*s", $widths[$i], '-';
                }
                elsif ($OPTS{raw}) {
                    printf " %*.1f", $widths[$i], $p;
                }
                else {
                    printf " %*.2f", $widths[$i], $p * 100;
                }
            }
            print "\n";
        }
    }
}



# grind_print_compact(): like grind_print(), but display a single perl
# in a compact form. Has an additional arg, $which_perl, which specifies
# which perl to display.
#
# Arguments are of the form:
#    $results->{benchmark_name}{perl_name}{field_name} = N
#    $averages->{perl_name}{field_name} = M
#    $perls = [ [ perl-exe, perl-label ], ... ]
#    $tests->{test_name}{desc => ..., ...}

sub grind_print_compact {
    my ($results, $averages, $which_perl, $perls, $tests, $order) = @_;


    # the width to display for each column.
    my $width = $OPTS{raw} ? 7 : 6;

    # Print standard header.
    grind_blurb($perls);

    print "\nResults for $perls->[$which_perl][1]\n\n";

    my @test_names = sorted_test_names($results, $order, $perls);

    # Dump the results for each test.

     my @fields = qw( Ir Dr Dw
                      COND IND
                      COND_m IND_m
                      Ir_m1 Dr_m1 Dw_m1
                      Ir_mm Dr_mm Dw_mm
                    );
    if ($OPTS{fields}) {
        @fields = grep exists $OPTS{fields}{$_}, @fields;
    }

    printf " %*s", $width, $_      for @fields;
    print "\n";
    printf " %*s", $width, '------' for @fields;
    print "\n";

    for my $test_name (@test_names) {
        my $doing_ave = ($test_name eq 'AVERAGE');
        my $res = $doing_ave ? $averages : $results->{$test_name};
        $res = $res->{$perls->[$which_perl][1]};

        for my $field (@fields) {
            my $p = $res->{$field};
            if (!defined $p) {
                printf " %*s", $width, '-';
            }
            elsif ($OPTS{raw}) {
                printf " %*.1f", $width, $p;
            }
            else {
                printf " %*.2f", $width, $p * 100;
            }

        }

        print "  $test_name\n";
    }
}


# do_selftest(): check that we can parse known cachegrind()
# output formats. If the output of cachegrind changes, add a *new*
# test here; keep the old tests to make sure we continue to parse
# old cachegrinds

sub do_selftest {

    my @tests = (
        'standard',
        <<'EOF',
==32350== Cachegrind, a cache and branch-prediction profiler
==32350== Copyright (C) 2002-2013, and GNU GPL'd, by Nicholas Nethercote et al.
==32350== Using Valgrind-3.9.0 and LibVEX; rerun with -h for copyright info
==32350== Command: perl5211o /tmp/uiS2gjdqe5 1
==32350== 
--32350-- warning: L3 cache found, using its data for the LL simulation.
==32350== 
==32350== I   refs:      1,124,055
==32350== I1  misses:        5,573
==32350== LLi misses:        3,338
==32350== I1  miss rate:      0.49%
==32350== LLi miss rate:      0.29%
==32350== 
==32350== D   refs:        404,275  (259,191 rd   + 145,084 wr)
==32350== D1  misses:        9,608  (  6,098 rd   +   3,510 wr)
==32350== LLd misses:        5,794  (  2,781 rd   +   3,013 wr)
==32350== D1  miss rate:       2.3% (    2.3%     +     2.4%  )
==32350== LLd miss rate:       1.4% (    1.0%     +     2.0%  )
==32350== 
==32350== LL refs:          15,181  ( 11,671 rd   +   3,510 wr)
==32350== LL misses:         9,132  (  6,119 rd   +   3,013 wr)
==32350== LL miss rate:        0.5% (    0.4%     +     2.0%  )
==32350== 
==32350== Branches:        202,372  (197,050 cond +   5,322 ind)
==32350== Mispredicts:      19,153  ( 17,742 cond +   1,411 ind)
==32350== Mispred rate:        9.4% (    9.0%     +    26.5%   )
EOF
        {
            COND    =>  197050,
            COND_m  =>   17742,
            Dr      =>  259191,
            Dr_m1   =>    6098,
            Dr_mm   =>    2781,
            Dw      =>  145084,
            Dw_m1   =>    3510,
            Dw_mm   =>    3013,
            IND     =>    5322,
            IND_m   =>    1411,
            Ir      => 1124055,
            Ir_m1   =>    5573,
            Ir_mm   =>    3338,
        },
    );

    for ('t', '.') {
        last if require "$_/test.pl";
    }
    plan(@tests / 3 * keys %VALID_FIELDS);

    while (@tests) {
        my $desc     = shift @tests;
        my $output   = shift @tests;
        my $expected = shift @tests;
        my $p = parse_cachegrind($output);
        for (sort keys %VALID_FIELDS) {
            is($p->{$_}, $expected->{$_}, "$desc, $_");
        }
    }
}