fnstcw in simple math functions?

fnstcw in simple math functions?

[Michael Vance]

Why is gcc emitting fnstcw instructions for this simple code? This
seems like a hopelessly conservative way of maintaining state. I tried
-ffast-math and got the same results.

int enclosed_mul( int x, float y )
{
    x = static_cast<int>( x * y );
    return x;
}

produces:

    .size   shortcut_mul__Fif,.Lfe2-shortcut_mul__Fif
    .align 4
.globl enclosed_mul__Fif
       .type	 enclosed_mul__Fif,@function
enclosed_mul__Fif:
	pushl %ebp
	movl %esp,%ebp
	subl $24,%esp
	fildl 8(%ebp)
	fmuls 12(%ebp)
	fnstcw -6(%ebp)
	movw -6(%ebp),%dx
	orw $3072,%dx
	movw %dx,-8(%ebp)
	fldcw -8(%ebp)
	fistpl -4(%ebp)
	movl -4(%ebp),%eax
	fldcw -6(%ebp)
	movl %ebp,%esp
	popl %ebp
	ret

This is using gcc-2.95.2 compiled as:

$ /usr/local/gcc-2.95.2/bin/g++ -O6 -fno-rtti -fno-exceptions -S
casting.cpp -o casting.s

m.

-- 
Programmer             "Ha ha." "Ha ha." "What are you laughing at?"
Loki Software                      "Just the horror of being alive."
http://lokigames.com/~briareos/                   - Tony Millionaire

Re: fnstcw in simple math functions?

[Zack Weinberg]

On Wed, Mar 15, 2000 at 02:39:51PM -0800, Michael Vance wrote:
> Why is gcc emitting fnstcw instructions for this simple code? This
> seems like a hopelessly conservative way of maintaining state. I tried
> -ffast-math and got the same results.
> 
> int enclosed_mul( int x, float y )
> {
>     x = static_cast<int>( x * y );
>     return x;
> }
...

This is because conversion from float to int is required to truncate,
but all other math operations are supposed to do round-to-nearest.

It might be interesting to see what kind of code we got if we
pretended that there was no way to put the 387 in truncate mode.  We'd
have to do more math, but fldcw/fnstcw is so slow that it'd probably
come out faster anyway.  Paging Jan Hubicka?

[I see glibc changes the control word in floor().  Bleah.]

zw