Re: [PATCH] Fix PR optimization/13653

[Gabriel Dos Reis <gdr at integrable-solutions dot net>]

Eric Botcazou <ebotcazou@libertysurf.fr> writes:

| Hi,
| 
| This is the long-standing PR for the POOMA regressions present on the 3.3 
| branch at -O2 -funroll-loops.  They are related to RTX_UNCHANGING_P (passing 
| -fno-schedule-insns2 make them disappear).
| 
| The problem is in the loop optimizer: if a giv happens to carry the /u flag, 
| is reduced but not replaceable, moves from the new register to the original 
| one with the /u flag are generated, thus breaking the semantics of 
| RTX_UNCHANGING_P.  This is further exacerbated here because the loop is 
| unrolled twice, so an unchanging read from the first loop can be intertwined 
| with an unchanging write from the second loop.
| 
| The proposed fix is to prevent givs that carry the /u flag and are not 
| replaceable from being reduced by the loop optimizer.  Bootstrapped/regtested 
| on i586-redhat-linux-gnu (3.3 branch).  OK for all branches?

This PR is reported to be already fixed on 3.4.x and only 3.3.x is
affected.  Please apply it to gcc-3_3-branch.

-- Gaby