MIPS 32-bit: Single perc. floating point constants in .rodata?

[Marshall B. Rogers]

Bump!

Would still be interested in an explanation if somebody has one :-P

Regards,
Marshall

On Mon, Jul 6, 2009 at 11:25 AM, Marshall B. Rogers <mbr@64.vg> wrote:
>
> Hi,
>
> I am in the process of rewriting some old code from ~1998 that was
> originally built with GCC (don't know the exact version number). The
> goal of the rewrite was to produce faster and more compact code, which
> has been an easy task so far considering how well GCC's optimization
> ability has progressed since then. Unexpectedly, my new program was
> larger than the original, and it turns out that this was due to 32-bit
> floating point constants being thrown into '.rodata', loaded from
> there, and then stored at their target. So,
>    lui     v0,0x809c
>    lwc1    $f0,8192(v0)
>    swc1    $f0,424(s0)
> Was being generated, whereas the original 1998 code had this:
>    lui     at,0x3f00
>    mtc1    at,$f4
>    swc1    $f4,424(s0)
> It's also important to note that the above could be simplified down to:
>    lui     at,0x3f00
>    sw      at,424(s0)
>
> Is there a switch that controls this behavior? If not, when did GCC
> stop behaving this (1998) way, and why? Potential 4-byte instruction
> overhead and 4-byte data overhead seems like a big waste to me!
>
> Regards,
> Marshall