[Bug c/31263] New: Misoptimization of constant function expressions

[Bug c/31263] New: Misoptimization of constant function expressions

[o dot mangold at gmx dot de]

I'm not sure, if this can be called a bug, but it is at least a really bad case
of poor optimization.

The following program calls the function 'Square' several times, either with
x=1000 or x=i*2-i-i+1000 (which is also 1000). The second version is executed
much FASTER. I see no reason, why this should be so. I tested it with gcc 4.1.1
and 4.1.2. The timings are more or less equal.

> gcc -O2 f_demo.c f_demo2.c -o f_demo
> time f_demo

real    0m1.537s
user    0m1.183s
sys     0m0.345s

> gcc -D VARIABLE_PAR -O2 f_demo.c f_demo2.c -o f_demo
> time f_demo

real    0m0.700s
user    0m0.368s
sys     0m0.329s

--- f_demo.c -----------------------------------------------------------
#include <stdlib.h>

double Square(double x);

#ifdef VARIABLE_PAR
#define PAR i*2-i-i+1000
#else
#define PAR 1000
#endif

int main()
{
  const int iSize=50000000;
  int i;
  double *pdA=malloc(iSize*sizeof(double));
  for(i=0;i<iSize;i++) {
    pdA[i]=Square(PAR);
  }
}

--- f_demo2.c ----------------------------------------------------
double Square(double x)
{
  return x*x;
}


-- 
           Summary: Misoptimization of constant function expressions
           Product: gcc
           Version: 4.1.2
            Status: UNCONFIRMED
          Severity: minor
          Priority: P3
         Component: c
        AssignedTo: unassigned at gcc dot gnu dot org
        ReportedBy: o dot mangold at gmx dot de
 GCC build triplet: i386-redhat-linux
  GCC host triplet: i386-pc-linux-gnu
GCC target triplet: i386-redhat-linux


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=31263

[Bug middle-end/31263] Misoptimization of constant function expressions

[pinskia at gcc dot gnu dot org]

------- Comment #1 from pinskia at gcc dot gnu dot org  2007-03-19 15:47 -------
On the trunk, there is no difference in the assembly produced for PPC-darwin.


-- 

pinskia at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
          Component|c                           |middle-end


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=31263

[Bug middle-end/31263] Misoptimization of constant function expressions

[rguenth at gcc dot gnu dot org]

------- Comment #2 from rguenth at gcc dot gnu dot org  2007-03-19 16:47 -------
On x86 different addressing modes and induction variables are used.  Good:

.L2:
        movq    %r12, 8(%rsp)
        addl    $1, %ebp
        movsd   8(%rsp), %xmm0
        call    Square
        movsd   %xmm0, (%rbx)
        addq    $8, %rbx
        cmpl    $50000000, %ebp
        jne     .L2

bad:

.L2:
        movq    %r12, 8(%rsp)
        movsd   8(%rsp), %xmm0
        call    Square 
        movsd   %xmm0, (%rbp,%rbx,8)
        addq    $1, %rbx
        cmpq    $50000000, %rbx
        jne     .L2

this is because we retain the complex expression until the last DOM run
after loop optimization (for 4.1.2 at least).


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=31263

[Bug middle-end/31263] Misoptimization of constant function expressions

[o dot mangold at gmx dot de]

------- Comment #3 from o dot mangold at gmx dot de  2007-03-20 09:21 -------
Maybe I should remark, that on my system the assembler output looks I little
different than yours.

Good:

.L2:
        pushl   %edi
        fildl   (%esp)
        addl    $4, %esp
        fstpl   (%esp)
        call    Square
        fstpl   (%esi,%ebx,8)
        addl    $1, %ebx
        cmpl    $50000000, %ebx
        jne     .L2

Bad:

.L2:
        movl    $0, (%esp)
        movl    $1083129856, 4(%esp)
        call    Square
        fstpl   -8(%esi,%ebx,8)
        addl    $1, %ebx
        cmpl    $50000001, %ebx
        jne     .L2


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=31263

[Bug target/31263] addressing modes are not selected correcly for x86 always

[pinskia at gcc dot gnu dot org]

------- Comment #4 from pinskia at gcc dot gnu dot org  2007-04-03 06:33 -------
To reproduce this on the trunk use "#define PAR i*2+1000-i-i" instead for the
complex one.

For PPC, VARIABLE_PAR is actually slower, as it tries to use (long unsigned
int) i * 8 as the index and PPC does not have that as an index mode.

And even the RTL optimizers are able to remove the i*2-i-i as we get:
((int) ((unsigned int) i + 1000) - i)
in final_cleanup so we were able to clean up one -i but not both on the tree
level.

The x86 issue is a target issue and I will report this other issue not saying
i*2+1000-i-i is 1000 as a differnet issue.


-- 

pinskia at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |pinskia at gcc dot gnu dot
                   |                            |org
          Component|middle-end                  |target
           Keywords|                            |missed-optimization
            Summary|Misoptimization of constant |addressing modes are not
                   |function expressions        |selected correcly for x86
                   |                            |always


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=31263