[Bug tree-optimization/38497] PRE missing a load PRE which causes a loop to have two BBs

[pinskia at gcc dot gnu.org via Gcc-bugs]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=38497

--- Comment #10 from Andrew Pinski  ---
CLANG/LLVM does this while ICC peels off the first iteration of the loop.

[Bug tree-optimization/38497] PRE missing a load PRE which causes a loop to have two BBs

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=38497

Richard Biener  changed:

   What|Removed |Added

   Last reconfirmed|2010-03-03 10:57:32 |2017-6-26

--- Comment #9 from Richard Biener  ---
I think the issue is we are not able to value-number both loads of *a to the
same value because DoHuffIteration clobbers *a.  So the *a in the block
leading to exit and the *a inserted in the latch do not have the same value.

I don't see how we can fix that easily inside PRE - we'd have to somehow
fix the value-numbering done during PHI translation to catch this case.

Iterating VN/PRE makes the value-numbering part work but hoisting still
fails because we look at ANTIC_IN & ~AVAIL_OUT but *a_6(D) is cleared
from ANTIC_OUT(3) as DoHuffIteration clobbers it.  So we really want
to look at ANTIC_OUT & ~AVAIL_OUT in do_hoist_insertion (ANTIC_IN is
a good approximation but it doesn't work here).  Of course ANTIC_OUT
we'd have to re-compute as we don't retain it.  There's the slight
complication that hoist insertion possibly inserts before the last
stmt though (that makes using ANTIC_IN a "fix" for that stmt possibly
clobbering the value).  Anyway, w/o fixing the value-numbering part
fixing hoisting is pointless of course.

[Bug tree-optimization/38497] PRE missing a load PRE which causes a loop to have two BBs

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=38497

--- Comment #8 from Andrew Pinski  ---
(In reply to Richard Biener from comment #5)
> now what is left is hoisting *a_2(D) to after the call.  But that's not PRE.

Yes I agree with this but hoisting was added to PRE and we still don't optimize
it ...

[Bug tree-optimization/38497] PRE missing a load PRE which causes a loop to have two BBs

[rguenth at gcc dot gnu dot org]

--- Comment #7 from rguenth at gcc dot gnu dot org  2010-05-05 10:30 ---
As *a_2(D) is ANTIC_OUT in bb3 (but not ANTIC_IN in bb3 because it dies in
there)
can we insert in bb3 instead of in bb5?


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=38497

[Bug tree-optimization/38497] PRE missing a load PRE which causes a loop to have two BBs

[rguenth at gcc dot gnu dot org]

--- Comment #3 from rguenth at gcc dot gnu dot org  2010-03-03 10:43 ---
This isn't PRE but code hoisting which we have a dup for.


-- 

rguenth at gcc dot gnu dot org changed:

   What|Removed |Added

 CC||stevenb dot gcc at gmail dot
   ||com


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=38497

[Bug tree-optimization/38497] PRE missing a load PRE which causes a loop to have two BBs

[steven at gcc dot gnu dot org]

--- Comment #4 from steven at gcc dot gnu dot org  2010-03-03 10:57 ---
I think pinskia means we could transform the test case of comment #0 to:

void DoHuffIteration(int);
int f(int *a)
{
  int i;
  int plaintextlen=*a;
  pretmp = plaintextlen;
  for(i = 0; i 1; i++)
 {
   DoHuffIteration(pretmp);
   pretmp = *a;
  }
  return pretmp - plaintextlen;
}

which makes this PRE, not hoisting.


-- 

steven at gcc dot gnu dot org changed:

   What|Removed |Added

 CC|stevenb dot gcc at gmail dot|
   |com |
 Status|UNCONFIRMED |NEW
 Ever Confirmed|0   |1
   Last reconfirmed|-00-00 00:00:00 |2010-03-03 10:57:32
   date||


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=38497

[Bug tree-optimization/38497] PRE missing a load PRE which causes a loop to have two BBs

[rguenth at gcc dot gnu dot org]

--- Comment #5 from rguenth at gcc dot gnu dot org  2010-03-03 11:03 ---
I'm not so sure.  We start with

bb 2:
  plaintextlen_3 = *a_2(D);

bb 3:
  # i_13 = PHI i_6(5), 0(2)
  D.2725_5 = *a_2(D);
  DoHuffIteration (D.2725_5);
  i_6 = i_13 + 1;
  if (i_6 = )
goto bb 5;
  else
goto bb 4;

bb 5:
  goto bb 3;

bb 4:
  D.2725_7 = *a_2(D);

so *a_2 (4) is antic-out in bb3 but dies in bb3 (thus it's not antic-in there).
*a_2 (3) is antic-in in bb5 and available in bb2.  Thus we insert in bb5:

bb 2:
  plaintextlen_3 = *a_2(D);

bb 3:
  # i_13 = PHI i_6(5), 0(2)
  # prephitmp.3_14 = PHI pretmp.2_1(5), plaintextlen_3(2)
  D.2725_5 = prephitmp.3_14;
  DoHuffIteration (D.2725_5);
  i_6 = i_13 + 1;
  if (i_6 = )
goto bb 5;
  else
goto bb 4;

bb 5:
  pretmp.2_1 = *a_2(D);
  goto bb 3;

bb 4:
  D.2725_7 = *a_2(D);

now what is left is hoisting *a_2(D) to after the call.  But that's not PRE.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=38497

[Bug tree-optimization/38497] PRE missing a load PRE which causes a loop to have two BBs

[steven at gcc dot gnu dot org]

--- Comment #6 from steven at gcc dot gnu dot org  2010-03-03 14:26 ---
Well, it is not hoisting, either.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=38497

[Bug tree-optimization/38497] PRE missing a load PRE which causes a loop to have two BBs

[pinskia at gcc dot gnu dot org]

--- Comment #2 from pinskia at gcc dot gnu dot org  2010-03-02 19:01 ---
Still happens as of today on the trunk.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=38497

[Bug tree-optimization/38497] PRE missing a load PRE which causes a loop to have two BBs

[pinskia at gcc dot gnu dot org]

--- Comment #1 from pinskia at gcc dot gnu dot org  2008-12-11 23:24 ---
This is a reduced testcase from The huffman byte benchmark.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=38497