[llvm-commits] [PATCH] [dragonegg] Use alias_pairs, fixes PR10587

[Duncan Sands]

Hi Peter,

>>> this is basically reverting commit 134356.  I would rather find another way of
>>> getting hold of these aliases if possible, since this way makes use of the LTO
>>> infrastructure and I was hoping to be able to not make any use of that one day.
>>> Any ideas?
>>
>> what do you think of doing this in a PLUGIN_ALL_IPA_PASSES_START callback
>> instead?
>
> Yes, that seems sensible, since that callback is called unconditionally
> before gcc empties the alias list.  New patch attached.

thanks for the patch.  I applied it with some modifications in commit 137021.
In particular I changed to the IPA end callback: since IPA passes seem to be
those that muck around with globals, it seemed best to do it after.  I moved
emission of unused globals to the same callback.

> Why are you avoiding the use of the LTO infrastructure?  It seems
> that you are aiming to support the use of gcc's LTO frontend with
> dragonegg, correct?

These are different issues.  If someone builds GCC without LTO support it would
be nice if the plugin still worked (and there seems to be no real reason why it
shouldn't).  If someone builds GCC with LTO support it would be good for the
plugin to play nicely with LTO; one such way is supporting GCC's LTO frontend.
But LTO shouldn't be required by the plugin.

Ciao, Duncan.