[Bug tree-optimization/87074] [8/9 Regression] Unroll and jam bug: O3 result differ from O2

[matz at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=87074

Michael Matz  changed:

   What|Removed |Added

   Keywords|needs-bisection |
 Status|NEW |RESOLVED
 Resolution|--- |FIXED
   Assignee|unassigned at gcc dot gnu.org  |matz at gcc dot gnu.org

--- Comment #7 from Michael Matz  ---
Fixed in trunk and gcc-8-branch

[Bug tree-optimization/87074] [8/9 Regression] Unroll and jam bug: O3 result differ from O2

[matz at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=87074

--- Comment #6 from Michael Matz  ---
Author: matz
Date: Sat Sep  1 17:33:45 2018
New Revision: 264030

URL: https://gcc.gnu.org/viewcvs?rev=264030=gcc=rev
Log:
Fix PR87074

Backport from mainline
PR tree-optimization/87074
* gimple-loop-jam.c (unroll_jam_possible_p): Check loop exit
PHIs for outer-loop uses.

testsuite/
* gcc.dg/pr87074.c: New test.

Added:
branches/gcc-8-branch/gcc/testsuite/gcc.dg/pr87074.c
Modified:
branches/gcc-8-branch/gcc/ChangeLog
branches/gcc-8-branch/gcc/gimple-loop-jam.c
branches/gcc-8-branch/gcc/testsuite/ChangeLog

[Bug tree-optimization/87074] [8/9 Regression] Unroll and jam bug: O3 result differ from O2

[matz at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=87074

--- Comment #5 from Michael Matz  ---
Author: matz
Date: Sat Sep  1 17:22:05 2018
New Revision: 264029

URL: https://gcc.gnu.org/viewcvs?rev=264029=gcc=rev
Log:
Fix PR87074

PR tree-optimization/87074
* gimple-loop-jam.c (unroll_jam_possible_p): Check loop exit
PHIs for outer-loop uses.

testsuite/
* gcc.dg/pr87074.c: New test.

Added:
trunk/gcc/testsuite/gcc.dg/pr87074.c
Modified:
trunk/gcc/ChangeLog
trunk/gcc/gimple-loop-jam.c
trunk/gcc/testsuite/ChangeLog

[Bug tree-optimization/87074] [8/9 Regression] Unroll and jam bug: O3 result differ from O2

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=87074

--- Comment #4 from Jakub Jelinek  ---
outer->header is:
  bb_4 (preds = {bb_3 bb_11 }, succs = {bb_5 })
  {
 [local count: 118111601]:
# f.10_31 = PHI 
# .MEM_34 = PHI <.MEM_18(D)(3), .MEM_21(11)>
# g_lsm.15_33 = PHI 
_75 = (unsigned int) g_lsm.15_33;
  }
the bb into which the inner loop exits is:
  bb_6 (preds = {bb_5 }, succs = {bb_11 bb_7 })
  {
 [local count: 118111601]:
# _11 = PHI <_8(5)>
_9 = (int) _11;
_12 = f.10_31 + 1;
if (_12 <= 10)
  goto ; [89.00%]
else
  goto ; [11.00%]
  }
and outer->latch is:
  bb_11 (preds = {bb_6 }, succs = {bb_4 })
  {
 [local count: 105119325]:
goto ; [100.00%]
  }
>From what I can see, gimple-loop-jam.c verifies using bb_prevents_fusion_p
these 3 bbs, but nothing checks the outer loop's IVs.
The problem is that a value computed by the inner loop (_8) is then used inside
of the outer loop and fed into the next inner loop too.

So, shall we punt if there are any phis in single_exit (loop)->dest (other than
virtual) (assuming here that we are in loop closed SSA form) and thus there are
any SSA_NAME uses from the inner loop to next outer loop's iteration?
Or are there perhaps any exceptions that can be handled, like if a value
computed by the inner loop is not used only after the outer loop exits?