On Thu, 28 May 2015, Ralf Baechle wrote:
> > The jump to the delay slot combined with the unusual register usage
> > convention taken here made it trickier than it would normally be to make a
> > fix that does not regress -- in terms of code size -- unaffected microMIPS
> > systems. I tried several versions and eventually I came up with this one
> > that I believe produces the best code in all cases, at the cost of these
> > #ifdefs. I hope they are acceptable.
>
> I think it's all a hint to rewrite the thing in a language that
> transparently handles the DADDIU issue. Such as C. Which would also
> make using a better algorithm easier.
Probably. One concern that bothers me is the ability of GCC to make
alternative entry points into frameless leaf functions.
Here we have `__strnlen_kernel_asm' that falls through to
`__strnlen_kernel_nocheck_asm'. That's a nice optimisation (we could
probably schedule that `move $v0, $a0' into its preceding delay slot too,
even though one might consider it hilarious to have a function's entry
point in a delay slot).
It would likely be lost in a conversion to C. But perhaps GCC can get
better, or maybe it already has? I haven't been tracking what's been
happening recently on that front.
What I have in mind is that given:
bar() { blah; }
foo() { blah_blah; bar(); }
in a single compilation unit, rather than making `foo' tail-jump to `bar'
GCC would inline `bar' into `foo' entirely and merely export an additional
`bar' entry point in the middle of `foo', where the original body of `bar'
begins.
Maciej
