------- Comment #6 from manu at gcc dot gnu dot org 2008-09-22 16:43 ------- Then, try to raise the issue in GCC and outline a clear plan. Otherwise, I assure you this is going to stay as-is for the next 5-10 years because nobody has a clear idea on how to tackle this in a way that pleases a majority. Also, because it has an easy workaround: use a casting to unsigned.
Some starting points: * Moving the warning to the middle-end is not going to avoid warning in this testcase unless VRP is enabled. For -O0 and -O1, we will keep warning. * Moving the warning to the middle-end means that we will warn/not warn depending on the ordering of passes, missed optimisations, etc. * Clang people claim that they can do a simple CCP pass very efficiently in the FE. Perhaps we could do the same. However, we are currently moving optimizations (such as folding) to later stages. In any case, VRP in the front-end is what would be needed for this testcase. Is that even possible? Can be done efficiently? The same issues apply to many warnings: -Wconversion, -Wuninitialized, -Wtype-limits. As Richard said: This is the usual trade-off of doing warnings from the frontends versus from the middle-end. One gives consistency and speed but frequent false positives, and the other gives more precision at the cost of consistent results and more compilation time. -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37591
