[Bug tree-optimization/22184] tree vectorizer depends on context

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=22184

Richard Guenther rguenth at gcc dot gnu.org changed:

   What|Removed |Added

 Blocks||53947

--- Comment #4 from Richard Guenther rguenth at gcc dot gnu.org 2012-07-13 
08:36:08 UTC ---
Link to vectorizer missed-optimization meta-bug.

[Bug tree-optimization/22184] tree vectorizer depends on context

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=22184

Richard Guenther rguenth at gcc dot gnu.org changed:

   What|Removed |Added

 Status|NEW |RESOLVED
 Resolution||FIXED

--- Comment #5 from Richard Guenther rguenth at gcc dot gnu.org 2012-07-13 
14:02:17 UTC ---
This is fixed, even if I disable VRP / jump-threading and substitute the info
from comment #3 in the code:

double MyFunc2 (int size)
{
  int len = size + 1;
  double Data[16] = {0};
if (len  3)
  ;
else if (len  0)
  ;
else
  return 0;

  for (int i=0; ilen; i++)
Data[i] = Data[i]=0 ? Data[i] : -Data[i];

  return Data[1];
}

[Bug tree-optimization/22184] tree vectorizer depends on context

[pinskia at gcc dot gnu dot org]

-- 
   What|Removed |Added

 CC||spop at gcc dot gnu dot org
   Last reconfirmed|2005-07-07 07:47:46 |2005-09-24 16:25:06
   date||


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=22184

[Bug tree-optimization/22184] tree vectorizer depends on context

[irar at il dot ibm dot com]

--- Additional Comments From irar at il dot ibm dot com  2005-07-07 07:47 
---
The problem occurs in decision whether the number of loop iterations is greater 
than zero. The (single) predecessor edge is checked for being EDGE_TRUE_VALUE 
or EDGE_FALSE_VALUE, and the corresponding predicate is used to make the 
decision. 

In the first case (single loop)  BB 0 contains predicate 'len0', its TRUE 
successor is BB 1, and the fallthru successor of BB 1 is BB 2 - the loop. The 
condition to check is 'len = 0', which is therefore simplified to FALSE. 

In the second case, however, the control flow is more complicated. The loop is 
in BB 6, its predecessor is BB 3, which has 2 predecessors: BB 5 (with 
predicate 'len  0'), and BB 2 - the first loop. The first loop is also guarded 
by 'len   0', but this information is not propagated.

-- 
   What|Removed |Added

 CC||rakdver at atrey dot karlin
   ||dot mff dot cuni dot cz
   Last reconfirmed|2005-06-25 18:28:42 |2005-07-07 07:47:46
   date||


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=22184

[Bug tree-optimization/22184] tree vectorizer depends on context

[rakdver at gcc dot gnu dot org]

--- Additional Comments From rakdver at gcc dot gnu dot org  2005-07-07 
10:52 ---
More precisely, the code when it comes to loop optimizer looks basically as

if (len  3)
  something;
else if (len  0)
  something_else;
else
  return;

for (i = 0; i  len; i++)
  whatever;

So indeed, len  0 on each path that reaches the loop, but it is not trivial to 
deduce.

One solution is to use the results of VRP here. However, VRP uses scev 
analysis, and adding yet another cycle into the current completely 
incomprehensible nest of dependences between # of iterations analysis and SCEV 
seems quite scary to me (and there are other technical difficulties with this 
solution -- running VRP twice is bad from efficiency reasons, but keeping 
results of VRP valid is not entirely trivial and may prevent some 
optimizations).

The other possibility is to extend the current oracle-like approach (# of 
iterations analysis basically asks whether a given condition -- len  0 in this 
case -- is true, and answer is computed on-demand by traversing SSA and 
dominance tree) to handle also this case.  This would however need to have the 
ASSERT_EXPRs (so we either would need to insert them before loop optimizer, or 
keep them valid since VRP, neither of which seems to be good for performance 
and memory consumption).

-- 
   What|Removed |Added

 CC||dnovillo at redhat dot com


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=22184

[Bug tree-optimization/22184] tree vectorizer depends on context

[pinskia at gcc dot gnu dot org]

--- Additional Comments From pinskia at gcc dot gnu dot org  2005-06-25 
18:28 ---
Confirmed, I think this is a bug in scaler evolution though.

-- 
   What|Removed |Added

   Severity|normal  |enhancement
 Status|UNCONFIRMED |NEW
 Ever Confirmed||1
   Keywords||missed-optimization
   Last reconfirmed|-00-00 00:00:00 |2005-06-25 18:28:42
   date||


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=22184