On 03/01/2017 07:03 AM, Richard Biener wrote:
The following addresses a regression in uninit warnings that happens
because clobber stmts preclude the very simple-minded support we have
for memory. The patch fixes this by instead implementing uninit
warnings for memory properly, using the alias oracle walk_aliased_vdefs
helper.
The patch adds better limiting to that interface and fixes one
false positive in fixed-value.c. Two other false positives are
fixed by the wide-int.h patch posted a few hours ago and a patch
to genemit from Jakub.
Bootstrap and regtest running on x86_64-unknown-linux-gnu with those
prerequesites included.
One issue with the patch is duplicate warnings as TREE_NO_WARNING
doesn't work very well on tcc_reference trees which are not
shared. A followup could use some sort of hash table to mitigate
this a bit. OTOH for maybe-uninit uses multiple locations may
be in need to be fixed to silence the warning. Another thing is
that we walk the function in random (BB) order and thus the
alias oracle walk limiting may result in warnings popping up
and going away in less predictable order (and also be reported
in odd order, like not all must-uninits first).
Comments? I realize this may introduce (a lot of?) false positives
quite late in the game, more aggressive limiting, like to 2
disambiguations, could solve the testcase in the PR and give up
most of the times while preserving the non-walking case of the
old code.
I'm all for it. Even if it triggers some false positives, improved
detection of uninit memory at compile time is a big win in my mind.
jeff
