Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
Reverse subregister saved loops in register usage info collector; NFC
On AMDGPU we have 70 register classes, so iterating over all 70 each time and exiting is costly on the CPU, this flips the loop around so that it loops over the 70 register classes first, and exits without doing the inner loop if needed. On my test just starting radv this takes RegUsageInfoCollector::runOnMachineFunction from 6.0% of total time to 2.7% of total time, and reduces the startup from 2.24s to 2.19s Patch by David Airlie! Differential Revision: https://reviews.llvm.org/D48582 llvm-svn: 340993
- Loading branch information