https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64946
--- Comment #22 from rguenther at suse dot de <rguenther at suse dot de> --- On Wed, 8 Jun 2016, shiva0217 at gmail dot com wrote: > https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64946 > > --- Comment #21 from Shiva Chen <shiva0217 at gmail dot com> --- > Hi, Richard > > On following example > > int a = ABS_EXPR (b); > int c = a - 1; > > c will get it's range base on a which is ~[INT_MIN+1, -1] No, a has [0, INT_MAX] and thus c gets [-1, INT_MAX-1] currently. > Even if we transfer to > > int a = (int) ABSU_EXPR (b); > int c = a - 1; > > Therefore, c still have to consider INT_MIN. Only now we have to consider INT_MIN. This is why we should do the transform to ABSU_EXPR only when it will likely help or when it doesn't change things. Like for short types. Thus short s; int a = (int) s; int b = ABS_EXPR (a); int c = b - 1; can be transformed to unsigned short bp = ABSU_EXPR (s); int b = (int) bp; int c = b - 1; and VRP analysis result will be unchanged but vectorization is now happy and ABSU_EXPR does not invoke undefined behavior when s == SHORT_MIN.
