[Bug optimization/2806] if's are not combined together

[law at redhat dot com]

------- Additional Comments From law at redhat dot com  2004-03-05 03:53 -------
Subject: Re:  if's are not combined together 

In message <20040304030605.4965.qmail@sources.redhat.com>, "pinskia at gcc dot 
gnu dot org" writes:
 >
 >------- Additional Comments From pinskia at gcc dot gnu dot org  2004-03-04 0
 >3:06 -------
 >We are now missing one jump threading for some reason:
 >(From .DOM3):
 >{
 ><bb 0>:
 >  if (eax_1 == 65535) goto <L0>; else goto <L6>;
 >
 ><L0>:;
 >  if (edx_2 == 1) goto <L1>; else goto <L4>;
 >
 ><L1>:;
 >  abarney[0] = 5;
 >  abarney[1] = 6;
 >
 ><L4>:;
 >  edx_3 = edx_2 - 1;
 >  if (edx_3 == 0) goto <L5>; else goto <L6>;
 >
 ><L5>:;
 >  afred[0] = 2;
 >
 ><L6>:;
 >  return;
OK.  edx_3 is a single use variable and we can (in theory) propagate it into
the conditional and adjust the conditional.  ie

if (edx_3 == 0)

can be turned into

  if (edx_2 == 1)

Once that's done, then the jump threader can thread through that block.

I've got a patch which does that as a straightforward extension to
tree-ssa-forwprop.  I'm still evaluating whether or not I want to risk
introducing it now or wait until after we merge.

On a positive note, with that improvement to tree-ssa-forwprop tree-ssa
generates noticeably better code than the mainline GCC sources.


jeff



-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=2806