[Bug optimization/14851] New: suboptimal fp division

[Bug optimization/14851] New: suboptimal fp division

[uros at kss-loka dot si]

Floating point division in current CVS gcc expands x/x into (1/x)*x. This testcase:

double test(double x)
{
        return x/x;
}

is with gcc 3.5.0 20040405 (experimental) with -O2 -ffast-math compiled to:
test:
        pushl   %ebp
        movl    %esp, %ebp
        fldl    8(%ebp)
        fld1
        fdiv    %st(1), %st
        popl    %ebp
        fmulp   %st, %st(1)
        ret

And with gcc 3.2 20020903 (Red Hat Linux 8.0 3.2-7):
test:
        pushl   %ebp
        movl    %esp, %ebp
        fldl    8(%ebp)
        fdiv    %st(0), %st
        leave
        ret

-- 
           Summary: suboptimal fp division
           Product: gcc
           Version: 3.5.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P2
         Component: optimization
        AssignedTo: unassigned at gcc dot gnu dot org
        ReportedBy: uros at kss-loka dot si
                CC: gcc-bugs at gcc dot gnu dot org
 GCC build triplet: i686-pc-linux-gnu
  GCC host triplet: i686-pc-linux-gnu
GCC target triplet: i686-pc-linux-gnu


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=14851

[Bug optimization/14851] [3.5 Regression] suboptimal fp division with -ffast-math

[pinskia at gcc dot gnu dot org]

------- Additional Comments From pinskia at gcc dot gnu dot org  2004-04-05 15:05 -------
Confirmed, a regression from 3.4.0.

-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
     Ever Confirmed|                            |1
           Keywords|                            |pessimizes-code
      Known to fail|                            |3.5.0
      Known to work|                            |3.4.0
   Last reconfirmed|0000-00-00 00:00:00         |2004-04-05 15:05:59
               date|                            |
            Summary|suboptimal fp division      |[3.5 Regression] suboptimal
                   |                            |fp division with -ffast-math
   Target Milestone|---                         |3.5.0


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=14851

[Bug rtl-optimization/14851] [4.0 Regression] suboptimal fp division with -ffast-math

[uros at gcc dot gnu dot org]

------- Additional Comments From uros at gcc dot gnu dot org  2004-09-17 06:08 -------
This is caused by following code in expr.c, around line 7676:

--cut here--
    case RDIV_EXPR:
      /* Emit a/b as a*(1/b).  Later we may manage CSE the reciprocal saving
         expensive divide.  If not, combine will rebuild the original
         computation.  */
      if (flag_unsafe_math_optimizations && optimize && !optimize_size
	  && TREE_CODE (type) == REAL_TYPE
	  && !real_onep (TREE_OPERAND (exp, 0)))
        return expand_expr (build2 (MULT_EXPR, type, TREE_OPERAND (exp, 0),
				    build2 (RDIV_EXPR, type,
					    build_real (type, dconst1),
					    TREE_OPERAND (exp, 1))),
			    target, tmode, modifier);
--cut here--

It is obvious, that in current 4.0, combine is ignoring this construct.

BTW: gcc-3.3 has quite weird behaviour with testcases like the one above, when
numerator is constant. When dividing integer-like numerators (for example
123532.0 / x), combine converts expression back to original one, and for real
numerators (4241.432 / x), no conversion occurs.

I will try to fix mainline.

-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |ASSIGNED
      Known to fail|4.0                         |4.0.0
   Last reconfirmed|2004-07-09 03:49:33         |2004-09-17 06:08:21
               date|                            |


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=14851

[Bug rtl-optimization/14851] [4.0 Regression] suboptimal fp division with -ffast-math

[uros at gcc dot gnu dot org]

------- Additional Comments From uros at gcc dot gnu dot org  2004-09-17 06:09 -------
Taking bug.

-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
         AssignedTo|unassigned at gcc dot gnu   |uros at kss-loka dot si
                   |dot org                     |


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=14851

[Bug rtl-optimization/14851] [4.0 Regression] suboptimal fp division with -ffast-math

[cvs-commit at gcc dot gnu dot org]

------- Additional Comments From cvs-commit at gcc dot gnu dot org  2004-09-21 05:11 -------
Subject: Bug 14851

CVSROOT:	/cvs/gcc
Module name:	gcc
Changes by:	uros@gcc.gnu.org	2004-09-21 05:10:55

Modified files:
	gcc            : ChangeLog rtlanal.c 

Log message:
	PR rtl-optimization/14851
	* rtlanal.c (commutative_operand_precedence): Calculate
	precedence of stripped constant.

Patches:
http://gcc.gnu.org/cgi-bin/cvsweb.cgi/gcc/gcc/ChangeLog.diff?cvsroot=gcc&r1=2.5543&r2=2.5544
http://gcc.gnu.org/cgi-bin/cvsweb.cgi/gcc/gcc/rtlanal.c.diff?cvsroot=gcc&r1=1.200&r2=1.201



-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=14851

[Bug rtl-optimization/14851] [4.0 Regression] suboptimal fp division with -ffast-math

[uros at kss-loka dot si]

------- Additional Comments From uros at kss-loka dot si  2004-09-21 05:17 -------
CVS commit in comment #4 is only partial fix for this PR. Please look at
http://gcc.gnu.org/ml/gcc-patches/2004-09/msg01928.html

There are still troubles with testcases:
double test (double x) {
        return 1.1 / x;
}
[A fix is proposed for above testcase]

and

double test (double x, double y) {
        return 5.1 / x + 5.2 / y;
}


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=14851

[Bug rtl-optimization/14851] [4.0 Regression] suboptimal fp division with -ffast-math

[uros at kss-loka dot si]

------- Additional Comments From uros at kss-loka dot si  2004-09-27 09:19 -------
A proposed fix is here:
http://gcc.gnu.org/ml/gcc-patches/2004-09/msg02193.html

By moving three-linked combine pass in front of two-linked pass, combiner has
considerably more freedom to simplify instructions.

-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=14851

[Bug rtl-optimization/14851] [4.0 Regression] suboptimal fp division with -ffast-math

[bonzini at gcc dot gnu dot org]

------- Additional Comments From bonzini at gcc dot gnu dot org  2004-10-04 07:51 -------
Adding this to 17652 since the pending patch was considered too risky for stage3
(http://gcc.gnu.org/ml/gcc-patches/2004-09/msg02241.html)

-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
OtherBugsDependingO|                            |17652
              nThis|                            |


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=14851

[Bug rtl-optimization/14851] [4.0 Regression] suboptimal fp division with -ffast-math

[giovannibajo at libero dot it]

------- Additional Comments From giovannibajo at libero dot it  2004-10-04 09:40 -------
OK, then we need to find another solution for 4.0.0, as this is a regression 
from 3.4.0.

-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=14851

[Bug rtl-optimization/14851] [4.0 Regression] suboptimal fp division with -ffast-math

[uros at kss-loka dot si]

------- Additional Comments From uros at kss-loka dot si  2004-10-05 05:57 -------
Roger suggested to move the a/b -> a * (1.0/b) transformation from RTL to
tree-ssa (http://gcc.gnu.org/ml/gcc-patches/2004-09/msg02017.html). In this
case, transformation could be enhanced to combine multiple divisions, as it is
explained in Agner Fog's "How to optimize for the Pentium family of
microprocessors", section 3.7.: a1/b1 + a2/b2 can be combined to (a1*b2 + a2*b1)
/ (b1*b2).

However, there is another bug exposed - when multiplying two constants from
constant pool, combiner combines two memory references into const_double.

Uros.

-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=14851

[Bug rtl-optimization/14851] [4.0 Regression] suboptimal fp division with -ffast-math

[pinskia at gcc dot gnu dot org]

------- Additional Comments From pinskia at gcc dot gnu dot org  2004-12-24 19:27 -------
The orginal bug was fixed, the next bug is really a regression from 2.95.3 for all 3.3, 3.4.0, and 4.0.0.
I opened a new bug for that, PR 19150.
So I am closing as fixed.

-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|ASSIGNED                    |RESOLVED
         Resolution|                            |FIXED


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=14851