IR protocol is not the only requirement for a direct call, and
the updating of both "has ir protocol" flag and the directly-
callable path was racy in both interpreter and JIT. In addition,
full builds and JIT that happen at threshold=0 in sync with
execution were sometimes continuing down the non-direct path using
the direct logic. This resulted in some call protocols running
twice, potentially messing up state for the direct logic.

The new flag is "canCallDirect", indicating that a transition to
the next tier has happened. Paths through normal non-direct logic
that might trigger a tier upgrade now re-check to see if that has
happened, re-branching back to the direct path.

Leaving this logic in place, commented, for the moment.

* Fixes #3544.

…uality

checks to eliminate the branch instr.  If we know we are comparing true to
true then we also know we can just perform a jump.  This has an added big
potential benefit that it should eliminate the next BB from the CFG entirely
since it will no longer be reachable.

If two directly comparable values are never equal then the branch is replaced
with a noop.  This will probably not help at all once compiled down to bytecode
since the JVM will figure out we have a bogus conditional but it makes me
feel better (and it was driving me crazy to see 'while(true)' generate
'b_false(true)').

…re possible

…od internals

* This should be better .. but let us see what Travis thinks.

* From truffle-head

…ported in docker.

This ensures that in the block call protocol scenario,
for lambdas as well as non-lambdas, all non-return control flow
(exceptions, breaks, nonlocal-returns) goes through the GEB.

Yikes .. this has other ramifications that I haven't worked through yet.

This reverts commit 11e3766.

…mode

# Conflicts:
#	tool/jt.rb