[LRA] liveness update not done after elimination??

[Vladimir Makarov]

On 12-10-04 8:53 AM, Steven Bosscher wrote:
> Hello Vlad,
>
> Is it intentional that DF_LR_IN and DF_LR_OUT are not updated after
> "Updating elimination of equiv for reg..."? I have some checking in
> place in process_bb_lives at the end of the function, and it triggers
> on the test case. (Checking code and test case is at the end of this
> e-mail.) It show:
>
> LIVEIN-mismatch in bb 5:
> 66, 69,
> 66, 69, 73
> 73
> t.c: In function 'get_mem_attrs':
> t.c:47:1: internal compiler error: in process_bb_lives, at lra-lives.c:866
>   }
>   ^
>
> And that's correct if you look at the LR_IN set of basic block 5 in
> the IRA  (.193r.ira) dump:
>
Ops.  I'll look at this.  As I remember, I thought about the problem.  
But I did not wanted to update live info because it is expensive and 
constraint equiv substitution by safe cases.  Somehow I added unsafe 
cases (containing pseudos) and even added the code to spill pseudos in 
substitutions when they start to conflict with others because of 
widening live ranges but forgot about live info. I'll try to fix it as 
soon as possible.