Proposal for Optimized Character Type Function

[Wilco Dijkstra]

Hi,

So if you do something like this in a loop:

sum += isupper (arr[i]);

the generated code looks as follows on AArch64:

.L5:
	ldrb	w1, [x2], 1
	ldrh	w1, [x4, x1, lsl 1]
	and	w1, w1, 256
	add	w0, w0, w1
	cmp	x3, x2
	bne	.L5

Ie. it already does exactly what you are suggesting.

A potential improvement might be to locate the ctype array faster
(perhaps at fixed TLS offset) and avoid using a function call. This would
reduce the overhead of ctype calls that aren't in tight loops.

Cheers,
Wilco