Re: [RFC PATCH] MIPS: optimise 32-bit do_div() with constant divisor

[Ralf Baechle <ralf@linux-mips.org>]

On Thu, Nov 06, 2014 at 04:23:18PM +0000, Mans Rullgard wrote:

> This is an adaptation of the optimised do_div() for ARM by
> Nicolas Pitre implementing division by a constant using a
> multiplication by the inverse.  Ideally, the compiler would
> do this internally as it does for 32-bit operands, but it
> doesn't.
> 
> This version of the code requires an assembler with support
> for the DSP ASE syntax since accessing the hi/lo registers
> sanely from inline asm is impossible without this.  Building
> for a CPU without this extension still works, however.
> 
> It also does not protect against the WAR hazards for the
> hi/lo registers present on CPUs prior to MIPS IV.
> 
> I have only tested it as far as booting and light use with
> the BUG_ON enabled wihtout encountering any issues.
> 
> The inverse computation code is a straight copy from ARM,
> so this should probably be moved to a shared location.

Can you explain why you need __div64_fls()?  There's __fls() which on
MIPS is carefully written to make use of the CLZ rsp. DCLZ instructions
where available; the fallback implementation is looking fairly similar
to your code.

MADD is named MAD on some older CPUs; yet other CPUs don't have it
at all.  I take it you tried to make GCC emit the instruction but it
doesn't?

  Ralf