== This week ==
* Applied patch for doing NEON high/low extraction using subregs. Ramana pointed out that we do the same thing for insertion, so I wrote a patch to handle that too. Both now merged into Linaro sources.
* Looked at ARM bootstrap problem on trunk. Turned out to be an aliasing problem. Submitted and applied patch.
* Reworked part of my SMS register-scheduling patch after feedback from Ayal. Submitted new version upstream.
* Got SPEC2006 running on the powerpc boxes and tested one part of my -fsched-pressure patch. Bit of a mixed bag. h264ref was one of the worst sufferers, which was a bit worrying. I think I'll need to make a third change too.
To recap, there are two pieces now:
1) Make -fsched-pressure honour the DFA 2) Make -fsched-pressure allow values that are live across a loop to be spilled.
I naively hoped that (1) would be OK on its own, but h264 shows that the current -fsched-pressure code is very conservative when it comes to large blocks. It only considers register deaths once there is a single remaining use; if there are two unscheduled uses, it assumes that the register remains live for the rest of the block.
So the problem that (1) was fixing was that -fsched-pressure was too optimistic in terms of what it could schedule in a cycle. But with that fixed, we seem to have too many sources of pessimism...
Richard