[PATCH] Don't upgrade global constructors when reading bitcode

[Duncan P. N. Exon Smith]

> On 2014-Aug-11, at 16:50, Reid Kleckner <rnk at google.com> wrote:
> 
>> IMO, writing to bitcode and reading back shouldn't change the IR.
>> 
>> The motivating problem I'm solving is serializing use-list order to
>> bitcode (PR5680).
> 
> The global_ctors auto-upgrade path is only one of many changes we make to IR when deserializing it, though, so by this argument we shouldn't do any auto-upgrade at all.

On the contrary, this as different from other auto-upgrades.  This code
is taking new IR and transforming it through bitcode serialization.  IMO,
"upgrade" is actually a misnomer here.  This is canonicalization, and
causes unnecessary differences in the IR depending on whether it's
serialized to bitcode or not.

As far as I can tell, all the other auto-upgrades are taking old IR and
upgrading it to new IR -- which is a different sort of thing.