http://gcc.gnu.org/bugzilla/show_bug.cgi?id=39326
--- Comment #43 from Steven Bosscher <steven at gcc dot gnu.org> 2013-03-09
14:57:52 UTC ---
The problem with combine is only "collateral damage" from what dse1 is
doing to this function. It's loading stored values into registers and
replacing re-loads with those registers, as in:
@@ -747831,6 +780399,7 @@
474783: r569578:SF=r569577:SF-r177936:SF
REG_DEAD r569577:SF
REG_DEAD r177936:SF
+ 605734: r662460:SF=r569578:SF
474784: [const(`t18571.26403'+0x5fe8)]=r569578:SF
REG_DEAD r569578:SF
474786: r177942:SF=[const(`t12303.26335'+0x9008)]
...
487110: [const(`t24716.26406'+0x2ff0)]=r578824:SF
REG_DEAD r578824:SF
- 487113: r578827:SF=[const(`t18571.26403'+0x5fe8)]
+ 487113: r578827:SF=r662460:SF
487114: [const(`t24716.26406'+0x2ff4)]=r578827:SF
REG_DEAD r578827:SF
The distance between the store and load is large, and that is what is
causing combine to blow up -- walking all the insns between 605734 and
487113 (about 500 of them, which appears to be typical). This also
increases register pressure, causing the poor IRA/LRA/postreload numbers.
Also not great: dse is loading the same stored value into multiple
registers, e.g.:
@@ -147272,6 +175190,8 @@
110394: r288307:SF=r288306:SF-r37078:SF
REG_DEAD r288306:SF
REG_DEAD r37078:SF
+ 601084: r657810:SF=r288307:SF
+ 601085: r657811:SF=r288307:SF
110395: [r288302:DI]=r288307:SF
REG_DEAD r288307:SF
but I'm guessing (and verifying now) that this is cleaned up by fwprop2.
