[Bug optimization/13674] ICE in reload_cse_simplify_operands, at postreload.c:378 on PPC64

[lucier at math dot purdue dot edu]

------- Additional Comments From lucier at math dot purdue dot edu  2004-01-15 23:32 -------
Subject: Re:  ICE in reload_cse_simplify_operands, at postreload.c:378 on PPC64

> A workaround is to use -fnew-ra.

Thank you for that suggestion.  Indeed, I could not find an example 
where -mcpu=G5 -fnew-ra either generated incorrect code or ICEd.

However, I would call -fnew-ra a partial workaround for the following 
reasons:

(a) Mainline gcc takes about 6 times as long to compile my code with 
-fnew-ra than without.

(b) For the most part, there is a performance regression of the 
compiled coded by about 50% with -fnew-ra than without.

(c) I believe that much of the new-ra development is sitting in its own 
branch, and it seems unlikely to me that this branch will be merged 
into mainline in time for the release of 3.4.

For these reasons I don't find -fnew-ra a complete or viable workaround.

It would be nice if Michael Matz were CC'ed on this discussion



-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=13674