Re: Fix reload ICE with define-split

[Ian Lance Taylor <iant at google dot com>]

Nathan Sidwell <nathan@codesourcery.com> writes:

> Roman Zippel wrote:
> 
> > An instruction can have multiple REG_INC notes and with your version
> > it would add a complete set of REG_INC notes for every previous
> > REG_INC note. IMO it's also a bit inconsequent to add REG_INC notes
> > for a register, which didn't have one before, only if there was a
> > REG_INC for a different register.
> > I'm fine with just recreating existing REG_INC notes or creating
> > REG_INC for all autoinc expressions (even new ones), I'm just not
> > sure about something inbetween.
> 
> I'm happy for an RTL maintainer to tell me what the right thing is :)

After define_split it needs to be the case that every autoincrement
has an associated REG_INC note.  Your second patch isn't right because
it adds all required REG_INC notes, but only if there was at least one
existing REG_INC note.

The question is: should we make the define_split responsible for
adding REG_INC notes when it introduces an autoincrement, or should we
slow down every define_split by searching for an RTX_AUTOINC code?  If
we choose the first option, then Roman's patch is correct.  If we
choose the second option, then your first patch is correct.

Existing backends already add REG_INC notes where required in a number
of different places.  So I am inclined to say that if a define_split
introduces an autoincrement, it should be responsible for adding a
REG_INC note (as is already done by sh.md).

So I'm going to approve Roman's patch, except that for increased
flexibility I think it should be modified to only create a REG_INC
note if the note does not already exist, since the backend may have
added the note itself.

Thanks.

Ian