On 05/14/14 03:06, Richard Biener wrote:
The following fixes pre/post-inc/dec gimplification of promoted
integer types. There is the issue with the way TYPE_OVERFLOW_UNDEFINED
is related to TYPE_OVERFLOW_WRAPS and the (non-)semantics of
-fno-strict-overflow.
In this case, with -On -fno-strict-overflow for a variable of
type short we have !TYPE_OVERFLOW_WRAPS _and_ !TYPE_OVERFLOW_UNDEFINED
(so we're in an "undefined" area). Which means that
!TYPE_OVERFLOW_UNDEFINED doesn't imply that overflow wraps.
Thus the gimplification has to play on the safe side and
always use an unsigned type unless the user specifies -fwrapv
(the flag with a proper semantic meaning).
That is, it seems to be the case that what predicate to use
(TYPE_OVERFLOW_WRAPS or TYPE_OVERFLOW_UNDEFINED, independent
on whether you invert it), depends on the use-case in a very
awkward (and error-prone) way.
Bootstrap and regtest pending on x86_64-unknown-linux-gnu, ok
if that succeeds (I expect to have to adjust some testcases)?
Thanks,
Richard.
2014-05-14 Richard Biener <rguent...@suse.de>
c-family/
* c-gimplify.c (c_gimplify_expr): Gimplify self-modify expressions
using unsigned arithmetic if overflow does not wrap instead of
if overflow is undefined.
* c-c++-common/torture/pr61184.c: New testcase.
Seems reasonable to me.
Jeff
