Version 4 (modified by bgamari, 4 years ago) (diff)


Say you have a program TestCase.hs in which a rewrite rule stopped firing at some point between GHC 7.10.1 and 7.10. It can often be useful to know which commit introduced the regression. Bisection is an efficient way of determining this, requiring at most log_2(N) commits to find the culprit among N commits.

This approach is especially appealing as git provides convenient support in the form of `git bisect`. git bisect coupled with a reliable test case and the script below (with appropriate modifications) turns the task of bisection into a relatively painless exercise.

Note: Bisecting revisions before the switch to submodules (i.e. more than a couple of months prior to the GHC 7.10.1 release) is quite difficult and is generally not worth the effort. The script below probably won't work in this regime.

To begin, it's best to minimize the work required to build GHC. For this setting BuildFlavour=quick in mk/ is recommended. Next download the script below and edit it to reflect your test-case. Now begin the bisection,

$ git bisect start
$ git bisect good ghc-7.10.1-release   # we know the testcase worked here
$ git bisect bad ghc-7.10.2-release    # but it fails here
$ git bisect run

This will run the script for each point in the bisection, skipping commits which are unbuildable. Hopefully this will end with a message informing you of the first bad commit. A log of this procedure will be place in $logs: $logs/all gives a nice high-level overview and the remaining files record each step of the bisection.

By default the script will clean the tree for every commit. While this is likely to give correct results, it performs a number of potentially redundant rebuilds. The process can be made faster by setting `ALWAYS_CLEAN=0", which will only clean the tree when a commit fails to build.



mkdir -p $logs
rev=$(git rev-parse HEAD)

function skip_commit() {
    exit 125

function log() {
    echo "$@" | tee -a $logs/all

function do_it() {
    log "Commit $rev: $step = $@"
    $@ 2>&1 | tee  $logs/$rev-$step.log
    log "Commit $rev: $step = $ret"
    return $ret

do_it submodules git submodule update || bad_commit
# We run `make` twice as sometimes it will spuriously fail with -j
if [ "x$ALWAYS_CLEAN" == "x0" ]; then
    # First try building without cleaning, if that fails then clean and try again
    do_it ghc1 make $make_opts || \
      do_it ghc2 make $make_opts || \
      do_it clean make clean && \
      do_it ghc3 make $make_opts || \
      do_it ghc4 make $make_opts || \
    do_it clean make clean || log "clean failed"
    do_it ghc1 make $make_opts || do_it ghc2 make $make_opts || skip_commit

# This is the actual testcase
# Note that this particular case depended upon the `text`
# library, which is checked out in $text
build="inplace/bin/ghc-stage2 $testcase -O -i$text -DMIN_VERSION_bytestring(x,y,z)=1 -I$text/include/"
do_it prebuild $build -fforce-recomp || skip_commit
touch $testcase
do_it build $build -ddump-rule-firings || skip_commit

# The test has succeeded if the rule fired 
if ! grep "Rule fired: TEXT literal" $logs/$rev-build.log ; then
    log "Commit $rev: failed"
    exit 1;
    log "Commit $rev: passed"
    exit 0;