[Bug middle-end/42574] [4.3/4.4/4.5/4.6 Regression] Address of global variable is calculated multiple times (missed CSE)

["steven at gcc dot gnu dot org" <gcc-bugzilla at gcc dot gnu dot org>]

------- Comment #8 from steven at gcc dot gnu dot org  2010-04-14 20:49 -------
Created an attachment (id=20379)
 --> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=20379&action=view)
Classic GCSE, resurrected (with some improvements)

Updated patch for trunk r158281.

Bootstrapped and tested on ia64-unknown-linux-gnu (with this new old pass
enabled at -O2 or higher with an incremental patch).

Also tested the effect on size using CSiBE with various different options.
Baseline options are "-Os -mcpu=arm7tdmi -fno-short-enums", other options are
"-fpic", "-mthumb", and "-fpic -mthumb":

                     trunk   patched  diff    ratio
baseline             3542232 3537080  5152    99.8546
baseline+fpic        3792475 3782747  9728    99.7435
baseline+mthumb      2693008 2690942  2066    99.9233
baseline+fpic+mthumb 2840739 2836169  4570    99.8391

So in all cases there is a code size reduction, but it's not very much. All in
all I'm not sure if it's worth the cost.

I'm also still working on integrating this with the hoisting pass. That would
make this pass easier to justify (it'd be almost free, HOIST already computes
everything).


-- 

steven at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
  Attachment #19446|0                           |1
        is obsolete|                            |


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=42574