[Bug tree-optimization/79578] Unnecessary instructions in generated code

2017-02-23 Thread law at gcc dot gnu.org 
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=79578

--- Comment #8 from Jeffrey A. Law  ---
Author: law
Date: Thu Feb 23 21:43:03 2017
New Revision: 245688

URL: https://gcc.gnu.org/viewcvs?rev=245688=gcc=rev
Log:
PR tree-optimization/79578
* tree-ssa-dse.c (clear_bytes_written_by): Use OEP_ADDRESS_OF
in call to operand_equal_p.

Modified:
trunk/gcc/ChangeLog
trunk/gcc/tree-ssa-dse.c

[Bug tree-optimization/79578] Unnecessary instructions in generated code

2017-02-22 Thread law at redhat dot com 
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=79578

Jeffrey A. Law  changed:

   What|Removed |Added

 Status|NEW |RESOLVED
 Resolution|--- |FIXED

--- Comment #7 from Jeffrey A. Law  ---
Fixed on the trunk.

[Bug tree-optimization/79578] Unnecessary instructions in generated code

2017-02-22 Thread law at gcc dot gnu.org 
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=79578

--- Comment #6 from Jeffrey A. Law  ---
Author: law
Date: Thu Feb 23 05:47:43 2017
New Revision: 245675

URL: https://gcc.gnu.org/viewcvs?rev=245675=gcc=rev
Log:
PR tree-optimization/79578
* tree-ssa-dse.c (clear_bytes_written_by): Use operand_equal_p
to compare base operands.

PR tree-optimization/79578
* g++.dg/tree-ssa/ssa-dse-3.C: New test.

Added:
trunk/gcc/testsuite/g++.dg/tree-ssa/ssa-dse-3.C
Modified:
trunk/gcc/ChangeLog
trunk/gcc/testsuite/ChangeLog
trunk/gcc/tree-ssa-dse.c

[Bug tree-optimization/79578] Unnecessary instructions in generated code

2017-02-17 Thread law at redhat dot com 
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=79578

Jeffrey A. Law  changed:

   What|Removed |Added

 CC||law at redhat dot com

--- Comment #5 from Jeffrey A. Law  ---
So at the first DSE pass the IL looks like:

;;   basic block 2, loop depth 0
;;pred:   ENTRY
  if (b_3(D) != 0B)
goto ; [0.00%]
  else
goto ; [0.00%]
;;succ:   3
;;4

;;   basic block 3, loop depth 0
;;pred:   2
  # .MEM_9 = VDEF <.MEM_5(D)>
  MEM[(struct A *)b_3(D)] = {};
;;succ:   4

;;   basic block 4, loop depth 0
;;pred:   3
;;2
  # .MEM_2 = PHI <.MEM_9(3), .MEM_5(D)(2)>
  # .MEM_12 = VDEF <.MEM_2>
  MEM[(struct A *)b_3(D)].a = 1;
  # .MEM_13 = VDEF <.MEM_12>
  MEM[(struct A *)b_3(D)].b = 2;
  # VUSE <.MEM_13>
  return b_3(D);
;;succ:   EXIT

The initialization looks like a fully dead store to me.  And it is.  The reason
it's not caught is we currently require pointer equality on the address
expression.  Using operand_equal_p instead results in the initialization being
eliminated by DSE1.  I think the only question is whether or not to fix this
during stage4 or wait for gcc-8.  It's a 1-liner ;-)

Of course once the dead store is gone BB3 is pointless and we can eliminate the
test resulting in:

movq%rdi, %rax
movl$131073, (%rdi)
ret

[Bug tree-optimization/79578] Unnecessary instructions in generated code

2017-02-17 Thread redi at gcc dot gnu.org 
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=79578

--- Comment #4 from Jonathan Wakely  ---
Since the check comes from the front-end, and the DR says it's not needed,
maybe this belongs to component c++ not tree-optimization. If the front-end
didn't insert the check we wouldn't need to optimize it away.

[Bug tree-optimization/79578] Unnecessary instructions in generated code

2017-02-17 Thread redi at gcc dot gnu.org 
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=79578

--- Comment #3 from Jonathan Wakely  ---
(In reply to Maxim Egorushkin from comment #0)
> 2. The code assigns through that pointer, so the pointer must be valid.
> Therefore there is no need to test the result of new for 0.

In any case, we fixed the standard so that it's undefined to pass null to the
placement new operator, so the check should never happen now, without needing
to see the origin of the pointer or whether it is subsequently derferenced.

http://www.open-std.org/jtc1/sc22/wg21/docs/cwg_defects.html#1748

(See also PR 35878 comment 2)

[Bug tree-optimization/79578] Unnecessary instructions in generated code

2017-02-17 Thread rguenth at gcc dot gnu.org 
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=79578

Richard Biener  changed:

   What|Removed |Added

   Keywords||missed-optimization
 Target||x86_64-*-*, i?86-*-*
 Status|UNCONFIRMED |NEW
   Last reconfirmed||2017-02-17
 CC||law at gcc dot gnu.org
  Component|c++ |tree-optimization
 Ever confirmed|0   |1

--- Comment #2 from Richard Biener  ---
Note that appearantly b is allowed to be NULL.

Confirmed for the partial dead code elimination issue.  tree DSE sees

   [100.00%]:
  if (b_3(D) != 0B)
goto ; [0.00%]
  else
goto ; [0.00%]

   [0.00%]:
  MEM[(struct A *)b_3(D)] = {};

   [0.00%]:
  MEM[(struct A *)b_3(D)].a = 1;
  MEM[(struct A *)b_3(D)].b = 2;
  return b_3(D);

and after store-merging we have (no further DSE):

   [100.00%]:
  if (b_2(D) != 0B)
goto ; [73.26%]
  else
goto ; [26.74%]

   [73.26%]:
  MEM[(struct A *)b_2(D)] = {};

   [100.00%]:
  MEM[(short unsigned int *)b_2(D)] = 131073;
  return b_2(D);

Jeff -- wasn't your "partial DSE" supposed to help here?  Ah, it was to
trim partial stores rather than full stores followed by partial ones
making it dead?