aco/live_var_analysis: misc improvements
Most of the patches have a slightly positive effect on compile times. For a randomly selected, more or less average shader the live var analysis is about ~15% faster with this series. The first 2 patches have the biggest impact.
I also tried:
- using a separate memory arena for the temporary live vars and
- set
propagate_on_container_copy_assignment
to false and - implement an explicit copy-assignment operator for IDSet which doesn't de-allocate the blocks
- using Temp instead of tempId for the IDSets (this was way worse!) and
- implement lower_bound() and upper_bound() in order be able access SGPRs vs VGPRs fast
- implement move-constructor and -assignment for
memory_resource
to free memory inreindex_vars()
(it segfaults, no idea how this works)