2017-04-10 12:13 GMT+02:00 Paul Gofman gofmanp@gmail.com:
On 04/09/2017 08:40 PM, Matteo Bruni wrote:
It looks to me as the CLIT section in the bytecode you attached contains 16 as the constants count. Not sure why the print below doesn't match that.
Either way, it doesn't matter much. Even if immediate counts non-multiple-of-4 can actually exist we could (and I argue, should) bump that up to the next multiple, zeroing the extra components.
Oh, it must be due to the code for updating constant table size underdiscussion does not respect scalar op flag, and 18 constants here reveals that (as the constant used in the instruction causing actual size update is using just 1 constant for all 3 components, not 3). It is not good by itself, in particular, it can break index wrap logic. For the same index wrap logic we can't just bump to next multiple of 4, if the cases where the count is not the multiple actually exist (I am not sure anymore that they do). So I will fix that update for scalar op case at some point, do some more testing and either send the fix or change the constant components count to 4, with a fail on preshader creation if it is not the case.
Ah, yes, that makes sense. Agreed on the follow up change.