Re: bootstrapping with profiling feedback?

[Nix <nix at esperi dot demon dot co dot uk>]

On Thu, 25 Jul 2002, Jeff Garzik spake:
> I'm assuming an implementation would look something like:
> 
> build stage1 using platform compiler
> build stage2 using stage1 and -fprofile-arcs
> build stage3 using stage2 and -fbranch-probabilities
> build stage4 using stage3 and -fbranch-probabilities
> compare stage 3 and 4 .o files
> 
> Maybe this could be called "make profilestrap"?  ;-)

I'd like that too; but I'd be just as happy to have a slightly different
target that builds stage1 with -fprofile-arcs -ftest-coverage, and
likewise for everything else (right up tot the libraries and testsuite),
so we can see how much of the (machine-independent and relevant
-dependent) parts of the compiler is actually being tested by a
bootstrap&test run.

(We might have to unconditionally turn off one or all of these flags for
the appropriate parts of profile.c and/or libgcc2.c themselves :) a small
loss, we'll *know* they're being tested...)


Pie-in-the-sky, I dream of a target that just runs the testsuite with a
compiler with test-coverage turned on, but builds libraries and things
without that flag, so we can tell how much of the compiler the testsuite
alone is testing; the whole-lot results might be perturbed by changes in
the compiler, but if we are testing part of the compiler via the
testsuite, it's reasonable to suppose that it'll stay tested :)

-- 
`There's something satisfying about killing JWZ over and over again.'
                                        -- 1i, personal communication