tCMS/git-hooks/pre-commit

#!/bin/sh

# Tidy staged Perl, judge it against .perlcriticrc, then run the tests, before
# any of it lands.
#
# There is a .perltidyrc, and a tree stays tidy to it only if something makes it.
# Left to people, it does not: tidying a file you are changing three lines of
# buries the change in a hundred lines of reformatting, so the reasonable thing
# is to skip it, and the tree drifts until a mass tidy is the only way back.
# One commit's worth at a time is small enough that nobody has to weigh it up.
#
# perlcritic is here for the other half of that.  Nothing else in tCMS runs the
# profile, so this is where an objection is about the diff in front of you.
#
# There is one profile, .perlcriticrc, and which one that is was decided when
# the floor was chosen: the packaging-perl skill asks which perl you
# are targeting and copies the matching profile to that name. So the hook needs
# no opinion about versions -- the file it reads is already the right one, and
# changing your floor means replacing that file rather than editing this.
#
# Install it: cp git-hooks/pre-commit .git/hooks/
#
# A file qualifies three ways:
#   - a .pm, .pl or .t extension
#   - a .PL extension, which is how a Makefile.PL gets covered
#   - executable with perl on its shebang line, which is how the extensionless
#     scripts in bin/ get covered
#
# .t is in there because test files are often not executable, and the shebang
# rule alone would then tidy whichever ones happen to be and let the rest drift.
#
# Both passes read the working tree rather than the index, so a file staged in
# part is judged whole. That is deliberate: perltidy rewrites the file on disk
# and the hook stages what it wrote, so the working copy is what lands anyway,
# and two passes disagreeing about which bytes they looked at would be worse to
# own than this.
#
# Both run from the top of the working tree, which is where git runs a hook and
# the only place they work at all: the profile names .preferred_modules.ini and
# .pod_stopwords by relative path.

# Redirect output to stderr.
exec 1>&2

# ACMR so that a file deleted in this commit is not handed to perltidy, which
# would fail on a path that is no longer there.
staged=$(git diff --cached --name-only --diff-filter=ACMR)

lib_files=''
test_files=''
for file in $staged
do
    test -f "$file" || continue

    is_perl=''
    case "$file" in
        *.pm|*.pl|*.PL|*.t)
            is_perl=1
            ;;
    esac

    if [ -z "$is_perl" ] && test -x "$file" && head -n 1 "$file" | grep -Eq '^#!.*perl'
    then
        is_perl=1
    fi

    test -n "$is_perl" || continue

    # Whether a file is perl at all, and then which list it goes in.  One case
    # statement doing both would file t/00-load.t under whichever arm it reached
    # first, and the .t rule is above the path rule.
    case "$file" in
        t/*) test_files="$test_files $file" ;;
        *)   lib_files="$lib_files $file" ;;
    esac
done

# For the tidy pass, which does not care which list a file came from.  The
# emptiness tests below concatenate the two without a space instead: with one,
# "no perl staged at all" would be a string holding a space, which is not empty.
all_files="$lib_files $test_files"

# Quoted here, because a bare [ $all_files ] is a syntax error the moment there
# is more than one file -- which would stop the hook working in exactly the
# commits that need it most.  Unquoted below, because there the word splitting is
# what turns one string into an argument list.
if [ -z "$lib_files$test_files" ]
then
    :
elif command -v perltidy >/dev/null 2>&1
then
    echo "Auto-tidying perl changes..."

    # -bext='/' makes perltidy delete the backup it would otherwise leave
    # sitting next to the file as <name>.bak.
    perltidy -b -bext='/' $all_files || {
        echo "pre-commit: perltidy failed, leaving the commit alone."
        exit 1
    }

    git update-index --add $all_files
    echo "Done."
else
    echo "pre-commit: perltidy is not on PATH, skipping the tidy pass."
fi

# After the tidy pass and never before.  The tidy pass rewrites these files and
# stages what it wrote, so what critic has to judge is the bytes that will land
# rather than the ones you saved -- and if you ever turn [CodeLayout::
# RequireTidyCode] back on, running it first would fail the commit over the one
# thing the hook had already set out to fix by itself.
#
# Three branches rather than two conditions and one else.  Staging nothing the
# profile judges -- a commit of documentation, say -- is not the same thing as
# having no perlcritic to judge it with, and one else says the second whenever
# the first is true.
if [ -z "$lib_files$test_files" ]
then
    echo "Nothing staged that the profile judges."
elif command -v perlcritic >/dev/null 2>&1
then
    echo "Running perlcritic..."

    # Both passes run even when the first has already failed, so one commit
    # shows every objection rather than one list's worth at a time.
    critic_failed=''

    if [ -n "$lib_files" ]
    then
        perlcritic --profile .perlcriticrc $lib_files || critic_failed=1
    fi

    # The same profile for t/, less three policies.  A test is not a module:
    # nobody runs perldoc on it, so requiring POD of it asks for a page no
    # reader will ever open.  And a test sets %ENV and %INC, and writes its
    # fixtures, in a BEGIN block before the SUT loads, for the whole file: a
    # local would end with the block, and no compiled binary is made of a test.
    # Everything else in the profile still judges it, which is more than a
    # release does -- all_critic_ok() reads lib/ and not t/, so this hook is the
    # only thing that looks at a test file at all.
    if [ -n "$test_files" ]
    then
        perlcritic --profile .perlcriticrc --exclude Documentation::RequirePod --exclude Variables::RequireLocalizedPunctuationVars --exclude CompileTime $test_files || critic_failed=1
    fi

    if [ -n "$critic_failed" ]
    then
        echo
        echo "pre-commit: perlcritic refused the above.  Either fix them, or say"
        echo "why with a '## no critic (Policy::Name)' and the reason beside it."
        echo "To commit without these checks at all: git commit --no-verify"
        exit 1
    fi

    echo "Done."
else
    # Skipped rather than fatal, for the same reason the tidy pass is: a machine
    # without perlcritic can still commit.
    echo "pre-commit: perlcritic is not on PATH, skipping the critic pass."
fi

# Last, because the tests read the files that the tidy pass wrote.  Whatever is
# staged, and not only Perl: a change to a share/ file or a fixture breaks a
# test as surely as a change to a module does.  -l for lib/.
#
# Without the GIT_ variables that git sets for this hook.  It sets
# GIT_INDEX_FILE, and GIT_DIR too in a worktree, and git obeys them over -C.  So
# a test that runs git in a repository of its own would add, commit and write
# config in this one.  In a subshell, because the git commands of this hook need
# them: under git commit -a, GIT_INDEX_FILE names the index being committed.
run_tests() {
    (
        unset $(env | sed -n 's/^\(GIT_[A-Za-z0-9_]*\)=.*/\1/p')
        prove -lm -j8 "$@"
    ) || {
        echo "pre-commit: the tests above failed.  To see why, run each failing file"
        echo "on its own:  prove -lv t/<file>.t"
        echo "To commit without these checks at all: git commit --no-verify"
        exit 1
    }
}

test -d t || exit 0

# Without tests-covering, every test, because nothing here knows which tests a
# change can reach.
if ! command -v tests-covering >/dev/null 2>&1
then
    echo "Running every test, because tests-covering is not on PATH; install Perl::Tests::Covering..."
    run_tests t/
    exit 0
fi

# tests-covering chooses the tests that ran a line the diff changes, from the
# records that git-hooks/post-commit keeps up to date.  It asks
# .tests-covering-map.pl about each file that no test loads, and the map names
# the files that no test reads, such as dist.ini.  --unexplained all runs every
# test for a file that neither can place, such as a fixture.  --full-index,
# because it matches each file to its records by the blob id on the index line
# of the diff, and an abbreviated id matches nothing.
tests=$(git diff --cached --full-index | tests-covering --diff --unexplained all) || exit 1
if [ -z "$tests" ]
then
    echo "No test ran a line that this commit changes."
    exit 0
fi

echo "Running the tests that ran what this commit changes..."
run_tests $tests