[Bug tree-optimization/99793] missed optimization for dead code elimination at -Os, -O2 and -O3 (vs. -O1)

rguenth at gcc dot gnu.org gcc-bugzilla@gcc.gnu.org
Mon Mar 29 07:35:56 GMT 2021


https://gcc.gnu.org/bugzilla/show_bug.cgi?id=99793

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
     Ever confirmed|0                           |1
           Assignee|unassigned at gcc dot gnu.org      |rguenth at gcc dot gnu.org
             Status|UNCONFIRMED                 |ASSIGNED
            Version|unknown                     |11.0
   Last reconfirmed|                            |2021-03-29
           Keywords|                            |missed-optimization

--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
So the difference is -fstrict-aliasing which causes us to not elide the
redundant store to 'd' which in turn makes us fail to promote 'd' read-only:

 Value numbering stmt = d = d.2_2;
-Store matched earlier value, value numbering store vdefs to matching vuses.
-Setting value number of .MEM_9 to .MEM_8 (changed)
-Deleted redundant store d = d.2_2;
+No store match
+Value numbering store d to d.2_2
+Setting value number of .MEM_9 to .MEM_9 (changed)

and the issue is that we "optimize" part of the walking with recording
last_vuse and using that vuse to insert the preceeding load into the
hashtables but since the walking is different for redundant store
detection (no TBAA), it cannot walk that far (beyond the non-TBAA
aliasing store to *b) and thus it cannot find the hashtable entry of
the d.2_2 = d load.  The last-vuse trick is poor-mans "PRE" - I do remember
experimenting with inserting both last and original vuse exprs but I do not
remember the outcome.


More information about the Gcc-bugs mailing list