[Bug target/78438] [7 Regression] incorrect comparison optimization

2016-11-23 Thread segher at gcc dot gnu.org 
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=78438

--- Comment #7 from Segher Boessenkool  ---
Author: segher
Date: Wed Nov 23 14:33:13 2016
New Revision: 242757

URL: https://gcc.gnu.org/viewcvs?rev=242757&root=gcc&view=rev
Log:
combine: Convert subreg-of-lshiftrt to zero_extract properly (PR78390)

r242414, for PR77881, introduces some bugs (PR78390, PR78438, PR78477).
It all has the same root cause: that patch makes combine convert every
lowpart subreg of a logical shift right to a zero_extract.  This cannot
work at all if it is not a constant shift, and it has to be a bit more
careful exactly which bits it extracts.


PR target/77881
PR bootstrap/78390
PR target/78438
PR bootstrap/78477
* combine.c (make_compound_operation_int): Do not convert a subreg of
a non-constant logical shift right to a zero_extract.  Handle the case
where some zero bits have been shifted into the range covered by that
subreg.

Modified:
trunk/gcc/ChangeLog
trunk/gcc/combine.c

[Bug target/78438] [7 Regression] incorrect comparison optimization

2016-11-21 Thread segher at gcc dot gnu.org 
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=78438

--- Comment #6 from Segher Boessenkool  ---
n/m, too tired I guess, please ignore comment 5.

[Bug target/78438] [7 Regression] incorrect comparison optimization

2016-11-21 Thread segher at gcc dot gnu.org 
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=78438

--- Comment #5 from Segher Boessenkool  ---
Why is this wrong?  This isn't the same reg 94 (it is reused):

Failed to match this instruction:
(set (mem/c:QI (symbol_ref:DI ("a") [flags 0x2]  )
[0 a+0 S1 A8])
(subreg:QI (ashiftrt:SI (reg:SI 93 [ b ])
(const_int 26 [0x1a])) 0))
Successfully matched this instruction:
(set (reg:SI 94)
(ashiftrt:SI (reg:SI 93 [ b ])
(const_int 26 [0x1a])))
Successfully matched this instruction:
(set (mem/c:QI (symbol_ref:DI ("a") [flags 0x2]  )
[0 a+0 S1 A8])
(subreg:QI (reg:SI 94) 0))

The low 8 bits of  r93 a>> 26  are exactly what we need to store afaics?

[Bug target/78438] [7 Regression] incorrect comparison optimization

2016-11-21 Thread law at redhat dot com 
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=78438

Jeffrey A. Law  changed:

   What|Removed |Added

   Priority|P3  |P1
 CC||law at redhat dot com

[Bug target/78438] [7 Regression] incorrect comparison optimization

2016-11-21 Thread jakub at gcc dot gnu.org 
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=78438

Jakub Jelinek  changed:

   What|Removed |Added

 Status|ASSIGNED|NEW

--- Comment #4 from Jakub Jelinek  ---
Oops, #c3 was meant for PR78436.

[Bug target/78438] [7 Regression] incorrect comparison optimization

2016-11-21 Thread jakub at gcc dot gnu.org 
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=78438

Jakub Jelinek  changed:

   What|Removed |Added

 Status|NEW |ASSIGNED

--- Comment #3 from Jakub Jelinek  ---
The over-sized bitfield isn't really needed for this, making it more severe:
struct S {
  long int : 23;
  long int a : 24;
  long int b : 10;
  long int c : 24;
  signed char d : 8;
} s;

__attribute__((noinline, noclone)) void
foo ()
{
  s.d = 0;
  s.c = -1193165L;
}

int
main ()
{
  foo ();
  if (s.d != 0)
__builtin_abort ();
  return 0;
}

[Bug target/78438] [7 Regression] incorrect comparison optimization

2016-11-21 Thread jakub at gcc dot gnu.org 
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=78438

--- Comment #2 from Jakub Jelinek  ---
Before combine we have:
(insn 6 5 7 2 (parallel [
(set (reg:SI 92)
(ashiftrt:SI (reg:SI 93 [ b ])
(const_int 11 [0xb])))
(clobber (reg:CC 17 flags))
]) "pr78438.c":7 563 {*ashrsi3_1}
 (expr_list:REG_DEAD (reg:SI 93 [ b ])
(expr_list:REG_UNUSED (reg:CC 17 flags)
(expr_list:REG_EQUAL (ashiftrt:SI (mem/c:SI (symbol_ref:DI ("b")
[flags 0x2]  ) [1 b+0 S4 A32])
(const_int 11 [0xb]))
(nil)
(insn 7 6 8 2 (parallel [
(set (reg:HI 94)
(lshiftrt:HI (subreg:HI (reg:SI 92) 0)
(const_int 15 [0xf])))
(clobber (reg:CC 17 flags))
]) "pr78438.c":7 572 {*lshrhi3_1}
 (expr_list:REG_DEAD (reg:SI 92)
(expr_list:REG_UNUSED (reg:CC 17 flags)
(nil
combine turns that into non-equivalent:
(insn 7 6 8 2 (parallel [
(set (reg:SI 94)
(ashiftrt:SI (reg:SI 93 [ b ])
(const_int 26 [0x1a])))
(clobber (reg:CC 17 flags))
]) "pr78438.c":7 563 {*ashrsi3_1}
 (expr_list:REG_UNUSED (reg:CC 17 flags)
(expr_list:REG_DEAD (reg:SI 93 [ b ])
(nil

[Bug target/78438] [7 Regression] incorrect comparison optimization

2016-11-21 Thread jakub at gcc dot gnu.org 
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=78438

Jakub Jelinek  changed:

   What|Removed |Added

 Status|UNCONFIRMED |NEW
   Last reconfirmed||2016-11-21
 CC||jakub at gcc dot gnu.org
   Target Milestone|--- |7.0
Summary|incorrect comparison|[7 Regression] incorrect
   |optimization|comparison optimization
 Ever confirmed|0   |1

--- Comment #1 from Jakub Jelinek  ---
Started with r242414.
Simplified testcase:
char a = 0;
int b = 197412621;

__attribute__ ((noinline, noclone))
void foo ()
{
  a = 0 > (short) (b >> 11);
}

int
main ()
{
  asm volatile ("" : : : "memory");
  foo ();
  if (a != 0)
__builtin_abort ();
  return 0;
}