Lennart Sorensen wrote:
There are also a couple of new constant load instructions "movw" and "movt" which make loading 16 and 32 bit constants more efficient.I believe the answer when armhf was originally defined is: Yes. Main reason as far as I understand it is that ARMv7 has thumb2, which is what armhf uses by default, which means the code size is smaller while having the same execution speed as normal arm instructions, but the smaller code size means less cache usage, so effectively your instruction cache is made larger and more effective, which certainly will help performance.
The question I haven't seen really answered is just how much difference this makes in practice. Anyone fancy loading debian armhf and raspbian on the same hardware and doing some benchmarking?
The version of armv6 used on the raspberry pi (which seems to be the only armv6 device anyone cares about) has vfpv2. Afaict vfpv3_d16 adds a constant load instruction, a couple of fixed point conversion instructions and makes exception support non-mandatory. Altogether not much difference.I am not sure if ARMv6 has the same VFP3-D16 as ARMv7, but I think it does.