[tree-ssa] moving phi nodes?

[tree-ssa] moving phi nodes?

[Richard Henderson]

One jump threading case that we're currently missing, sorely, 
is of the form:

int foo(int p, int q)
{
  int i = 0;
  if (p)
    i = q;
  if (i != 0)
    bar ();
  return i;
}

From the t28.dce2 dump:

  # BLOCK 0
  # PRED: ENTRY (fallthru,exec)
  if (p_3 != 0) goto <Uc910>; else goto <Uc984>;
  # SUCC: 2 (false,exec) 1 (true,exec)

  # BLOCK 1
  # PRED: 0 (true,exec)
<Uc910>:;
  # SUCC: 2 (fallthru,exec)

  # BLOCK 2
  # PRED: 0 (false,exec) 1 (fallthru,exec)
  # i_1 = PHI <0(0), q_4(1)>;
<Uc984>:;
  if (i_1 != 0) goto <Uc9f8>; else goto <Uca6c>;
  # SUCC: 4 (false,exec) 3 (true,exec)

  # BLOCK 3
  # PRED: 2 (true,exec)
<Uc9f8>:;
  bar ();
  # SUCC: 4 (fallthru,exec)

  # BLOCK 4
  # PRED: 2 (false,exec) 3 (fallthru,exec)
<Uca6c>:;
  return i_1;
  # SUCC: EXIT

We really should be threading the edge BB0->BB4, since it is known
that I == 0 when incoming to block 2.  Except if we do this, we need
to move the PHI node for I:

  # BLOCK 0
  # PRED: ENTRY (fallthru,exec)
  if (p_3 != 0) goto <Uc910>; else goto <Uca6c>;
  # SUCC: 4 (false,exec) 1 (true,exec)

  # BLOCK 1
  # PRED: 0 (true,exec)
<Uc910>:;
  # SUCC: 2 (fallthru,exec)

  # BLOCK 2
  # PRED: 1 (fallthru,exec)
<Uc984>:;
  if (q_4 != 0) goto <Uc9f8>; else goto <Uca6c>;
  # SUCC: 4 (false,exec) 3 (true,exec)

  # BLOCK 3
  # PRED: 2 (true,exec)
<Uc9f8>:;
  bar ();
  # SUCC: 4 (fallthru,exec)

  # BLOCK 4
  # PRED: 0 (false,exec) 2 (false,exec) 3 (fallthru,exec)
  # i_1 = PHI <0(0), 0(2), q_4(3)>;
<Uca6c>:;
  return i_1;
  # SUCC: EXIT

Moving the PHI node is non-trivial.  Anyone know of a way to do it that
doesn't require just dropping all the SSA names and starting over?  Do
we have any other gimple pass planned or desired that would take care 
of this as a side effect?

Incidentally, this optimization is performed by the rtl .14.bypass 
gcse pass.


r~

Re: [tree-ssa] moving phi nodes?

[law]

In message <20031116062517.GA2111@redhat.com>, Richard Henderson writes:
 >One jump threading case that we're currently missing, sorely, 
 >is of the form:
 >
 >int foo(int p, int q)
 >{
 >  int i = 0;
 >  if (p)
 >    i = q;
 >  if (i != 0)
 >    bar ();
 >  return i;
 >}
[ ... ]
 >Moving the PHI node is non-trivial.  Anyone know of a way to do it that
 >doesn't require just dropping all the SSA names and starting over?  Do
 >we have any other gimple pass planned or desired that would take care 
 >of this as a side effect?
 >
 >Incidentally, this optimization is performed by the rtl .14.bypass 
 >gcse pass.
I would expect this to be caught by the jump threading improvements I'm
working on.  Maybe not the first couple iterations, but it's the kind of
thing I do expect to handle.



Jeff

Re: [tree-ssa] moving phi nodes?

[law]

In message <20031116062517.GA2111@redhat.com>, Richard Henderson writes:
 >One jump threading case that we're currently missing, sorely, 
 >is of the form:
 >
 >int foo(int p, int q)
 >{
 >  int i = 0;
 >  if (p)
 >    i = q;
 >  if (i != 0)
 >    bar ();
 >  return i;
 >}
 >
 >From the t28.dce2 dump:
[ ... ]
FWIW, I just checked in an improvement to the dominator optimizer which 
should optimize this case.

jeff