[Bug tree-optimization/48092] New: associative property of sqrt

[Bug tree-optimization/48092] New: associative property of sqrt

[vincenzo.innocente at cern dot ch]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=48092

           Summary: associative property of sqrt
           Product: gcc
           Version: 4.6.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: tree-optimization
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: vincenzo.innocente@cern.ch


Is there any reason why (with -Ofast or -ffast-math) associative properties of
sqrt are not exploited as for instance those of division?

examples
division (ok)

float div1(float a, float x, float y) {
   return a/x/y;
   0:   f3 0f 59 ca             mulss  %xmm2,%xmm1
   4:   f3 0f 5e c1             divss  %xmm1,%xmm0
}

sqrt

float sqrt1(float a, float x, float y) {
   return a*std::sqrt(x)*std::sqrt(y);
  10:   f3 0f 51 d2             sqrtss %xmm2,%xmm2
  14:   f3 0f 51 c9             sqrtss %xmm1,%xmm1
  18:   f3 0f 59 ca             mulss  %xmm2,%xmm1
  1c:   f3 0f 59 c8             mulss  %xmm0,%xmm1
}
  20:   0f 28 c1                movaps %xmm1,%xmm0

and


float rsqrt1(float a, float x, float y) {
   return a/std::sqrt(x)/std::sqrt(y);
  30:   f3 0f 51 c9             sqrtss %xmm1,%xmm1
  34:   f3 0f 51 d2             sqrtss %xmm2,%xmm2
  38:   f3 0f 59 d1             mulss  %xmm1,%xmm2
  3c:   f3 0f 5e c2             divss  %xmm2,%xmm0
}

in this second case I would have at least expected the use of 
"rsqrtss" to take precedence above the associative property of "div"
emitting the same code as below

float rsqrt2(float a, float x, float y) {
   return a/sqrtf(x*y);
  70:   f3 0f 59 ca             mulss  %xmm2,%xmm1
  74:   f3 0f 52 d9             rsqrtss %xmm1,%xmm3
  78:   f3 0f 59 cb             mulss  %xmm3,%xmm1
  7c:   f3 0f 59 cb             mulss  %xmm3,%xmm1
  80:   f3 0f 59 1d 00 00 00    mulss  0(%rip),%xmm3        # 88 <rsqrt2(float,
float, float)+0x18>
  87:   00 
                        84: R_X86_64_PC32       .LC1+0xfffffffffffffffc
  88:   f3 0f 58 0d 00 00 00    addss  0(%rip),%xmm1        # 90 <rsqrt2(float,
float, float)+0x20>
  8f:   00 
                        8c: R_X86_64_PC32       .LC0+0xfffffffffffffffc
  90:   f3 0f 59 cb             mulss  %xmm3,%xmm1
  94:   f3 0f 59 c8             mulss  %xmm0,%xmm1
}
  98:   0f 28 c1                movaps %xmm1,%xmm0

[Bug rtl-optimization/48092] missing optimization (associative property of sqrt)

[vincenzo.innocente at cern dot ch]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=48092

vincenzo Innocente <vincenzo.innocente at cern dot ch> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
          Component|tree-optimization           |rtl-optimization
            Version|4.6.0                       |4.7.0
            Summary|associative property of     |missing optimization
                   |sqrt                        |(associative property of
                   |                            |sqrt)

--- Comment #1 from vincenzo Innocente <vincenzo.innocente at cern dot ch> 2011-08-05 08:16:17 UTC ---
it is still not optimized in 4.7

[Bug tree-optimization/48092] associative property of builtins is not exploited on GIMPLE

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=48092

Richard Guenther <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2011.08.05 08:33:59
                 CC|                            |rguenth at gcc dot gnu.org
          Component|rtl-optimization            |tree-optimization
            Summary|missing optimization        |associative property of
                   |(associative property of    |builtins is not exploited
                   |sqrt)                       |on GIMPLE
     Ever Confirmed|0                           |1

--- Comment #2 from Richard Guenther <rguenth at gcc dot gnu.org> 2011-08-05 08:33:59 UTC ---
The reason is we handle this only from the frontend folder which doesn't see
builtins when using std::sqrt.  A C testcase like

double foo (double x, double y)
{
  return __builtin_sqrt (x) * __builtin_sqrt (y);
}

is optimized.

For C++ (thus, optimizing the above when wrapping the sqrt inside an
inline function) these optimizations have to be replicated at the 
GIMPLE level.  Which today means in, for example, tree-ssa-forwprop.c.

Same issue for std::exp and other associative builtins.

[Bug tree-optimization/48092] associative property of builtins is not exploited on GIMPLE

[vincenzo.innocente at cern dot ch]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=48092

--- Comment #3 from vincenzo Innocente <vincenzo.innocente at cern dot ch> 2011-09-08 10:01:48 UTC ---
btw even in C  with -Ofast
 a*exp(x)*exp(y) (same for sqrt) is NOT optimized. compare

double exp0(double x, double y) {
   return exp(x)*exp(y);
}

double exp1(double a, double x, double y) {
   return a*exp(x)*exp(y);
}

double exp02(double a, double x, double y) {
  return a*(exp(x)*exp(y));
}

[Bug tree-optimization/48092] associative property of builtins is not exploited on GIMPLE

[rguenther at suse dot de]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=48092

--- Comment #4 from rguenther at suse dot de <rguenther at suse dot de> 2011-09-08 10:50:46 UTC ---
On Thu, 8 Sep 2011, vincenzo.innocente at cern dot ch wrote:

> http://gcc.gnu.org/bugzilla/show_bug.cgi?id=48092
> 
> --- Comment #3 from vincenzo Innocente <vincenzo.innocente at cern dot ch> 2011-09-08 10:01:48 UTC ---
> btw even in C  with -Ofast
>  a*exp(x)*exp(y) (same for sqrt) is NOT optimized. compare
> 
> double exp0(double x, double y) {
>    return exp(x)*exp(y);
> }
> 
> double exp1(double a, double x, double y) {
>    return a*exp(x)*exp(y);
> }

Yes, that's because those simplification all are implemented in
fold-const.c.  Compare

double exp03(double a, double x, double y) {
  double tmp1 = exp(x);
  double tmp2 = exp(y);
      return a*(tmp1*tmp2);
}

and in fold-const.c we don't associate FP math.  tree-ssa-reassoc.c
re-associates exp1, but then the simplification is missing - see
my exp03 example.

[Bug tree-optimization/48092] associative property of builtins is not exploited on GIMPLE

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=48092

Andrew Pinski <pinskia at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|                            |missed-optimization
           Severity|normal                      |enhancement

[Bug tree-optimization/48092] associative property of builtins is not exploited on GIMPLE

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=48092

Andrew Pinski <pinskia at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
         Resolution|---                         |FIXED
   Target Milestone|---                         |6.0
             Status|NEW                         |RESOLVED

--- Comment #5 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
exp and sqrt were both fixed in GCC 6 by r6-2973-g354016403304