[llvm-commits] [llvm] r151466 - in /llvm/trunk: include/llvm/Analysis/Dominators.h lib/Analysis/InstructionSimplify.cpp

[Rafael Espíndola]

On 2 March 2012 17:44, Duncan Sands <baldrick at free.fr> wrote:
> Hi Rafael,
>
>> Don't call dominates on unreachable instructions.
>
> why not?  Is there really no sensible definition of domination in this case?

The textbook book one that I know is that (at least at the BB level),
an unreachable BB is dominated by anything. Chris points out that it
is handy to have an API where dominates(A, B) != dominates(B, A), and
that would be false if we used that definition.

Do you know of an alternative one that would fit better?

> For example, consider dominates(A, B).  Suppose the definition is: every path
> from the entry block to B passes through A.  Then if B is unreachable then
> dominates should return 'true'.  After all, every path from the entry block to
> B passes through A.  (If you don't agree, show me a path from the entry block
> to B that doesn't pass through A).  If B is reachable and A is unreachable,
> then dominates should return 'true' since there is a path from the entry
> block to B that doesn't pass through A.  But perhaps you have a different
> definition of domination in mind?

I had that one, and it looks like it makes most of the dominates uses
simpler, but I have not tried changing them all.

> Ciao, Duncan.
>
> PS: If you add an assertion about dominates(instruction,instruction) only
> being called for reachable instructions, then shouldn't you have one for
> dominates(basic block, basic block) too?  After all, if you think
> dominates(instruction,instruction) has no sensible meaning, probably
> dominates(basic block, basic block) has no sensible meaning either.

The functions are a bit different, that is why I used a different name
for dominates(inst, inst) in the first patch. While dominates(bb,bb)
is the classical textbook one, the (inst, inst) version is about
dominating uses.

Cheers,
Rafael