Arsen Arsenović wrote:
Hi,
On Saturday, 24 September 2022 06:24:56 CEST you wrote:
But when it's unsupported because ${tool}-dg-prune returns
"::unsupported::" the individual checks within the file still appear
in the results. I realise that in the target selector case, the test
is just skipped before doing anything, and in the prune case we
don't
know that it's unsupported until *after* testing it. But it seems to
me that ideally the individual checks would get "retroactively
skipped" if tool-dg-prune returns one of ::untested::,
::unresolved::, or ::unsupported::. Maybe that's not easy to do
though.
This actually looks like a bug in dg.exp to me. The output from
${tool}-dg-test is first collected, all at once, (line 696) and then
analyzed (lines 701-749). I would expect ${tool}-dg-prune to have an
opportunity to modify the text to be analyzed /before/ the analysis is
done, but the current code checks for messages first, /then/ calls
${tool}-dg-prune (and the framework's prune_warnings procedure) to
remove irrelevant output, then declares excess errors if the final
text is nonempty.
This is how I understood the output analysis routine: remove expected
lines first (i.e. those matched by dg-messages), then lines that
shouldn't be a FAIL but are noise or such instead (i.e. prune), which
allows ${tool}-dg-prune and dg-prune-output to be less careful about
what they remove (as they don't have to worry about removing diagnostics
that should be matched by dg-{error,warning}).
Changing ${tool}-dg-prune to execute earlier might be a bit too drastic,
because it could expose some overzealous pruning routines.
Running ${tool}-dg-prune earlier works both ways, allowing the callback
prune routine to also remove or modify any output that might confuse the
message-matching mechanism. Really, given the documented (such as it
is) feature for ${tool}-dg-prune to be able to declare the entire test
unsupported, I would expect it to filter the results before messages are
checked.
Overzealous pruning routines are bugs in the testsuites that have them,
since they would be very likely to hide /actual/ excess errors, thereby
causing the test for excess errors to PASS when it should FAIL.
On the other
hand, giving ${tool}-dg-test a way to terminate analysis early, before
all the output processing, could maybe constrained enough to not change
what existing code does (which is why I suggested it earlier).
What do you think?
Tcl provides a way to do that (I think) if you really want:
unsupported "freestanding libstdc++"
return -code return
...but this would be a bit of a hack.
-- Jacob
