[m-dev.] cannot_fail deconstruct -vs- goal_info determinism

[Fergus Henderson]

I've analyzed the bug that caused an internal compiler error when
extras/trailed_update/var.m was compiled with `-O4 --high-level-code'.

Deforestation was performance some optimizations, and then rerunning
mode analysis, but not rerunning determinims analysis.  One of the
deconstruction unifications was originally marked by mode analysis as
`can_fail' and hence by determinism analysis as `semidet'.  When mode
analysis was rerun by deforestation, the more precise insts lead to
the unification being marked as `cannot_fail'.  However, determinism
analysis was not rerun, so the determinism in the goal_info was still
`semidet'.  ml_unify_gen.m was checking that the `can_fail' field in
deconstruct unifications matched the determinism on the goal_info,
and this check failed, leading to the call to error/1.

It would be easy to modify ml_unify_gen.m so that it doesn't make that
assumption.  However, I'm not sure if that is the right solution.
There's three possible solutions, and I'm not sure which is best:

	- Document this invariant in hlds_goal.m, as an invariant that
	  must hold at all stages after determinism analysis, and modify
	  deforestation to preserve the invariant (e.g. by rerunning
	  determinism analysis after it reruns mode analysis).

	- Document this invariant in hlds_goal.m, as an invariant that
	  must hold when generating code, and add code in simplify.m to
	  ensure that the invariant holds.  (This would match what we do
	  for the invariant that the determinism match the instmap_delta
	  with regard to at_most_zero_solns/unreachable.)

	- Modify ml_unify_gen.m so that it doesn't depend on this
	  invariant, but instead allows the determinism in the
	  goal_info to be any conservative approximation.

Any suggestions as to which of these we should adopt?

-- 
Fergus Henderson <fjh at cs.mu.oz.au>  |  "I have always known that the pursuit
                                    |  of excellence is a lethal habit"
WWW: <http://www.cs.mu.oz.au/~fjh>  |     -- the last words of T. S. Garp.
--------------------------------------------------------------------------
mercury-developers mailing list
Post messages to:       mercury-developers at cs.mu.oz.au
Administrative Queries: owner-mercury-developers at cs.mu.oz.au
Subscriptions:          mercury-developers-request at cs.mu.oz.au
--------------------------------------------------------------------------