ia64 performance regression

[Steve Christiansen <smc at us dot ibm dot com>]

What is the purpose of the following code in loop.c (loop_regs_scan)?

  for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
    {
      regs->array[i].may_not_optimize = 1;
      regs->array[i].set_in_loop = 1;
    }

It seems to be regarding all hardware registers as being set within loops.
Why?

I'm asking the question because an ia64 global variable reference
looks like this:

(insn 52 48 54 (parallel[
            (set (reg/f:DI 360)
                (symbol_ref:DI ("p")))
            (clobber (reg:DI 361))
            (use (reg:DI 1 r1))
        ] ) -1 (nil)
    (expr_list:REG_EQUAL (symbol_ref:DI ("p"))
        (nil)))

r1 is the global data pointer.  The USE of r1 is causing a dependency
which is made non-invariant when loop_regs_scan sets "set_in_loop"
for r1.  By not hoisting global data references, we're leaving a lot of
ia64 performance on the table.

The problem was exposed by

http://gcc.gnu.org/ml/gcc-patches/2001-07/msg00908.html

Before that patch, dependencies introduced by USEs inside of PARALLELs were
simply ignored.

Here's a little C code that causes the problem to happen.
Compile with -O2 -S on ia64 and take a look at the .s file.

extern long *buf1, *buf2;
extern int n;

long *p, *q;

void doit ()
{
    int i;
    int m = n;
    long c = 12;

    p = buf1;
    q = buf2;

    for (i = 0; i < m; i++) {
        *(q++) = *(p++) + c;
    }
}

Any insights would be appreciated.

Thanks

Steve

-- 
Steve Christiansen <smc@us.ibm.com>
IBM Linux Technology Center
503-578-4177  IBM T/L: 775-4177
http://www.ibm.com/linux/ltc