Re: Emit more REG_EQUIV notes for function args (PR42235)

[Jeff Law <law@redhat.com>]

On 07/14/10 15:47, Bernd Schmidt wrote:
> On 07/14/2010 08:54 PM, Jeff Law wrote:
>    
>> In theory, given the REG_EQUIV note we ought to get an entry in
>> reg_equiv_mem, which the code I'm working on knows it can use instead of
>> shoving the pseudo into a stack slot.  Effectively, my code will
>> rematerialize the argument from the equivalent memory location
>> regardless of the number of uses.
>>      
> Reload can do that already, I think, but it's not prepared to handle a
> (zero_extend (mem)) inside a REG_EQUIV note.  That should be reasonably
> trivial to add if the reg is never set other than in the initializing insn.
>    
reload does, but it's code that can be problematical.  For example, 
replacement of an unallocated pseudo with its equivalent can trigger 
secondary reloads, spills, etc.

My code creates a new pseudo/allocno for the various ranges where the 
unallocated pseudo is live then asks IRA to color those split-range 
pseudos/allocnos.  What we end up presenting to reload is cleaner; 
though often they result in the same final code.  Handling this case 
falls out nicely from the infrastructure for splitting ranges of 
unallocated pseudos without equivalent forms (it's probably a couple 
dozen lines of new code).

jeff