[Mesa-dev] INTEL_DEBUG=shader_time scalar backend fixes

[Kenneth Graunke]

Welcome to the rabbit trail.  In order to fix Football Manager, I had to
rework INTEL_DEBUG=shader_time in the FS backend.  While doing that, I
hit two assertion failures.  After fixing that, I compared numbers.
I noticed that VS again accounted for 0 cycles.

Matt - could you take a look at brw_vec4_dead_code_elimination.cpp?
Since 5df88c2096281f (the rewrite to use live intervals), it's again
completely eating the atomic buffer offset initialization, resulting
in us using garbage data in the first register (of an mlen 2 message).
(Sounds like the same bug I fixed in d0575d98fc595dcc17706dc73d1eb4.)

With these patches, and the new vec4 DCE pass reverted, 10.3 vs my
branch produce basically the same numbers on an openarena timedemo.

Available in the 'football-v3' branch of my tree.