Re: [RFC] Track discriminators by instruction instead of by basic block

[Cary Coutant <ccoutant at google dot com>]

> I don't think that observation is correct (that tree-profile is after
> streaming in).

Yeah, you're right. It doesn't look like it's feasible to delay the
assignment of discriminators to instructions until after the IR is
streamed out. For the moment, I think we can live with losing the
discriminator information on a -flto compilation -- the important
thing is that they will be passed through a non-LTO initial
compilation so that samples from a profiling run can be mapped back to
the correct blocks. The profile-use compilation can use -flto, though,
as the discriminators will survive long enough for the frequency
information to be set correctly.

Nevertheless, I'll work on the fixing this before the next Stage 1
opens, and I'll try to make sure it doesn't bloat the external IR.
(Would one extra byte per instruction be considered excessive?)

Thanks!

-cary