On 8/16/2019 2:11 PM, Jeff King wrote: > On Thu, Aug 15, 2019 at 10:01:04PM -0400, Derrick Stolee wrote: > >> Here is today's test coverage report. > > Are the scripts you use to generate these available somewhere? The scripts I originally used are available in contrib/coverage-diff.sh. However, to allow for some more advanced workflows, I created the following C# project: https://github.com/derrickstolee/git-test-coverage/tree/master/TestCoverageReport The runs are done via the following Azure Pipelines build: https://dev.azure.com/git/git/_build/results?buildId=879 There, you can see how the code is built, tested, and the gcov output is collected in the TestCoverageReport tool to generate the .txt and .htm files. > I think it's useful to look at uncovered code, but I often struggle to > figure out whether the parts attached to my name are relevant. In > particular, I think two changes to the report format might help: > > 1. Include names alongside commit ids when listing uncovered lines. I > know that will end up with some overly-long lines, but it makes it > easy to grep for one's name to find relevant sections of the file > (as opposed to finding your name at the bottom and > cross-referencing with actual content lines). > > Seeing that an uncovered line is a BUG(), for example, makes it > easy to know that it's not really an interesting uncovered case in > the code. The HTML reports [1] have the following feature: click on a commit in the commit summary and it highlights the lines from that commit. [1] https://derrickstolee.github.io/git-test-coverage/reports/2019-08-15.htm However, the plain-text report _could_ add names. Maybe instead the report should just group the output by commit instead of branch. > 2. Include more context. Just taking a random example from this email: > >> builtin/rebase.c >> e191cc8b 129) strbuf_addstr(&buf, strategy_opts); > > We know what the uncovered line was trying to do, but more interesting > is likely the conditional that causes it to be uncovered. In this case > the surrounding code is: > > if (opts->ignore_whitespace) { > struct strbuf buf = STRBUF_INIT; > > if (strategy_opts) > strbuf_addstr(&buf, strategy_opts); > > strbuf_addstr(&buf, " --ignore-space-change"); > free(strategy_opts); > strategy_opts = strbuf_detach(&buf, NULL); > } > > even the usual 3 lines of diff context would make it a lot quicker to > understand what's going on (it only kicks in when multiple strategy > options are used). In this case, the additional context helps, but only if you expand by several lines. In other cases, the necessary context could be dozens of lines. Perhaps a deeper report could include something like "git format-patch --function-context" to naturally extend to the appropriate context. It may be better to have a "block context" for these conditionals. In such a case, it would be important to mark the "new" lines explicitly so the context lines don't become too noisy. > (As an aside, this code leaks the newly allocated buffer and leaves a > dangling pointer in opts->strategy_opts, but that's all orthogonal to > the uncovered line; I'll send a separate message to the original > author). > > > Anyway, I wonder if we could adjust the output of the script to make > reading it that way a bit easier. I'll think about this, for sure. When I have time to go deep into the report (not just the code I wrote) I look for lines that seem to be non-trivial, then go find their context by looking at the commit. Here is another example of why the HTML report is helpful: it will link to the commit diff view [2] or the exact line of the file [3]. (Note: the exact line link only works if we point to the branch that was tested, not the commit that made the change.) [2] https://github.com/git/git/commit/e191cc8b8080f63a0080031bf1276269c6bf42dd [3] https://github.com/git/git/blob/6cf2e4a04610efe882eb663cd36436b80b343fa0/builtin/rebase.c#L129 Thanks, -Stolee
