On 11/04/2016 12:18 PM, Pat Haugen wrote:
While working to get -fsched-pressure profitable on PowerPC, one of the things I noticed causing problems was the selection of pressure classes. Understanding that getting the computation of pressure classes correct for all targets in the machine independent code is going to be extremely hard (or impossible), I propose the following target hook which lets the target decide the set of pressure classes to use. I will follow this up with an implementation of the hook for the rs6000 backend, but am currently benchmarking variations of it to see which is best. Bootstrap/regtest on powerpc64le with no new regressions. Ok for trunk?
Yes. It is OK, Pat. Thanks for working on powerpc insn-scheduling issues.
Note that the ira.c changes are not as complicated as what it looks, I only added a few lines to invoke the target hook, the rest is just indentation changes to move the existing code into the else leg.
