[Bug middle-end/31249] pseudo-optimzation with sincos/cexpi

[dominiq at lps dot ens dot fr]

------- Comment #16 from dominiq at lps dot ens dot fr  2007-03-21 15:36 -------
> The recommended way is to post a message to gcc@gcc.gnu.org or
> gcc-patches@gcc.gnu.org

I'll follow your advice, but before I'ld like some feedback about what follows.

I have applied the following patch

--- gcc-4.3-20070316/gcc/tree-ssa-math-opts.c   Thu Mar  8 20:02:51 2007
+++ gcc-4.3-20070317/gcc/tree-ssa-math-opts.c   Tue Mar 20 17:21:16 2007
@@ -704,9 +704,7 @@
 gate_cse_sincos (void)
 {
   /* Make sure we have either sincos or cexp.  */
-  return (TARGET_HAS_SINCOS
-         || TARGET_C99_FUNCTIONS)
-        && optimize;
+  return TARGET_HAS_SINCOS && optimize;
 }

 struct tree_opt_pass pass_cse_sincos =

I have regtested it with no change in the reports for gcc, g++, gfortran, 
and objc. The timings before and after are

                              before                   after
                        -O0     -O1      %      -O0     -O1      %

g++-4 sincos_o.c        6.2     9.6     +55     6.3     5.6     -11
gfc sincos_o.f90        6.3     9.6     +52     6.4     5.5     -14

for the following C and Fotran tests:

[karma] bug/timing% cat sincos_o.c
#include <math.h>
#include <stdio.h>

int main()
{

  long    n = 1000000;
  long    i;
  double mo = -1.0;
  double pi = acos(mo);
  double sc = 0.0;
  double ss = 0.0;
  double  t = 0.0;
  double dt = pi/n;

  printf("%.17g \n", pi);
  printf("%.17g \n", dt);
  for (i=0; i< 40*n; i++) {
    sc += cos(t);
    ss += sin(t);
    t += dt;
  }
  printf("%.17g %.17g \n", sc, ss);
}

[karma] bug/timing% cat sincos_o.f90 
integer, parameter :: n=1000000
integer :: i
real(8) :: pi, ss, sc, t, dt
pi = acos(-1.0d0)
dt=pi/n
sc=0
ss=0
t=0
do i= 1, 40*n
  sc = sc + cos(t)
  ss = ss + sin(t)
  t = t+dt
end do
print *, sc, ss
end

So from the PPC Darwin point of view, everything is working as expected.

I have done a search on the regtest list based on

FAIL: gcc.dg/builtins-59.c scan-tree-dump __builtin_cexpi

assuming that platforms that do not pass it, are likely to have not
__builtin_cexpi, thus are exposed to the same bad optimization.
I have found the following list tested on a regular basis:

powerpc-apple-darwin8.5.0
hppa2.0w-hp-hpux11.11
v850-unknown-elf
sparc-unknown-elf
sh-unknown-elf
powerpc-unknown-eabisim
mips-unknown-elf
m32r-unknown-elf
m32c-unknown-elf
avr-unknown-none
frv-unknown-elf
arm-unknown-elf
cris-axis-elf
arm-none-eabi

As far as I can tell, only the first two are tested against g++ and 
gfortran.

Note that the list does not include powerpc64-apple-darwin8.8.0.
So it seems that it has __builtin_cexpi.

I don't know what will be the final decision about the proposed patch,
but there is no "emergency" since I can use it for my coming weekly builds.
I would prefer to have some feedback from the listed platforms before
seeing the patch applied.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=31249