https://gcc.gnu.org/bugzilla/show_bug.cgi?id=69857
--- Comment #6 from David Binderman <dcb314 at hotmail dot com> ---
(In reply to ktkachov from comment #5)
> We'd need a testcase that shows a regression resulting from this code not
> being run i.e. code that became worse after r197530 (or wrong code or an
> ICE) and is fixed by removing that "return false;".
> Then, a normal bootstrap and test should be enough and the change could go
> in as a regression fix at this stage. But finding such a testcase would not
> be easy, I suspect.
A step on the way to finding that test case would be to put a marker
into the code. Something like
else if (TARGET_ARM)
{
// Either this
fprintf( stderr, "dcb: Got to interesting place\n");
// or this
inform (0, "dcb: Got to interesting place");
return false;
Then compile a lot of code, looking for the "interesting place" comment.
This can be left unattended, so need not consume valuable developer time.
A simple grep through some build logs would do.
Once we've got example code that gets to that place, merely do a before'n'after
with and without the "return false" and see what happens. If there is
a regression, then we are done.
I suspect the "return false;" merely prevents us doing some ARM
specific optimisation, so removing the "return false" is likely IMHO
to make the code go faster.
Just an idea.
