[Bug tree-optimization/96351] missed opportunity to optimize out redundant loop

[amacleod at redhat dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96351

Andrew Macleod <amacleod at redhat dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |amacleod at redhat dot com

--- Comment #2 from Andrew Macleod <amacleod at redhat dot com> ---
I had to do some tweaking and manually inline it, but eventually I managed to
get:

 <bb 4> :
  pszTmp_11 = str_8(D) + 2;
  goto <bb 6>; [INV]

  <bb 5> :
  pszTmp_13 = pszTmp_5 + 2;

  <bb 6> :
  # pszTmp_5 = PHI <pszTmp_11(4), pszTmp_13(5)>
  _1 = *pszTmp_5;
  if (_1 != 0)
    goto <bb 5>; [INV]
  else
    goto <bb 7>; [INV]

  <bb 7> :
  # pszTmp_16 = PHI <pszTmp_5(6)>
  _2 = pszTmp_16 - str_8(D);
  _3 = _2 /[ex] 2;
  len_12 = (unsigned int) _3;

<bb 8> :
  # len_4 = PHI <0(2), len_12(7), 0(3)>
  if (len_4 == 0)
    goto <bb 9>; [INV]
  else
    goto <bb 10>; [INV]


Nothing we can do right now.  Its possible that the relational work in the next
release *might* be able to sort something out, but thats probably down the
road.

We'll be able to record that
a) pszTmp_11 > str_8(D)
b) pszTmp_13 > pszTmp_5
c) when we see 
# pszTmp_5 = PHI <pszTmp_11(4), pszTmp_13(5)>
we might possibly be able to deduce that 
d) pszTmp_5 > str_8(D),  which then leads to
e) pszTmp_16 > str_8(D).

which means we could know that _2 > 0.

The tricky bit will be knowing that not only is _2 > 0, but its actually [2,
+INF].  Initial plans for relations are not to track possible offsets, but it
could be follow up work.

If we can know that, then we'd be able to calculate len_12 = [1, +INF/2].

With a non-zero range for len_12, then it would be possible for the threader to
create 2 execution paths.. the zero and the non-zero path.

on the non-zero path , there would no longer need to be a reference to len_12,
and dead code could conceivably eliminate the enter else block.

I'd keep this around as a reference, but I don't see this happening any time
soon. If the threaders get a revamp next year... then maybe!