[Bug rtl-optimization/39077] [4.3/4.4/4.5 Regression] GCSE-optimization causes enormous binary size increase (~20 times !)

[steven at gcc dot gnu dot org]

------- Comment #13 from steven at gcc dot gnu dot org  2009-04-12 23:46 -------
The real bug is that somehow MEM_ATTRS are not shared anymore.  We have lots
and lots of exactly the same expression in the table, e.g.:

Index 3 (hash value 4232)
  (mem/s/f/c:SI (plus:SI (reg/f:SI 20 frame)
        (const_int -3828 [0xfffffffffffff10c])) [32 cpy.d+0 S4 A32])
Index 6 (hash value 4232)
  (mem/s/f/c:SI (plus:SI (reg/f:SI 20 frame)
        (const_int -3828 [0xfffffffffffff10c])) [32 cpy.d+0 S4 A32])
Index 10 (hash value 4232)
  (mem/s/f/c:SI (plus:SI (reg/f:SI 20 frame)
        (const_int -3828 [0xfffffffffffff10c])) [32 cpy.d+0 S4 A32])


but exp_equiv_p() thinks these are not equivalent, because the MEM_ATTRS
pointers are not the same.  We should have MEM_ATTRS(x)==MEM_ATTRS(y) for two
MEMs with the same memory attributes, but here the pointers are not the same. 
So we're allocating MEM_ATTRS somewhere without going via the table, or we're
adjusting MEM_ATTRS somewhere wvia an incorrect interface.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=39077