Currently we do: > switch (type()) { > case MIRType_Int32: > outputRange.wrapAroundToInt32(); > break; > case MIRType_Boolean: > outputRange.wrapAroundToBoolean(); > break; > case MIRType_None: > MOZ_CRASH("Asking for the range of an instruction with no value"); > default: > break; > } Though it should be possible to do: clampAroundInt32. So investigate and try to do that. That should improve our ranges we have as output range.

(In reply to Hannes Verschore [:h4writer] from comment #0) > Though it should be possible to do: clampAroundInt32. > > So investigate and try to do that. That should improve our ranges we have as > output range. The reason we do this is "Truncate analysis". The range is an upperbound and only can get decreased in passes. It is disallowed to increase the range again. Now that is what 'truncate analysis' does: [0; +inf], becomes [MIN_INT32; MAX_INT32]. As a result if an instruction can possibly get truncated, we need to use: "wrapAroundToInt32" instead of "clampAroundInt32". Since we cannot inflate the clamped range [0; MAX_INT32] to [MIN_INT32; MAX_INT32] anymore!