Replace the instruction with the scalar variant FCVT instruction.
FCVT{N,L} 8 cycles latency on the Cortex A57
FCVT has five cycle latency and slightly higher throughput
On the A72 all three of these instructions will have three cycle latency,
While FCVT{N,L} will have half the throughput.
This provides a decent speed up in pretty much everything that touches pair loadstores because in most cases they are just regular non-quantizing
float loadstores that happen.
Instead of having an "INS" instruction after every single instruction to duplicate the bottom 64bits in to the top 64bits of the register,
create a new FPR register cache type to track when a register's lower 64bits is supposed to be duplicated in to the high 64bits.
Not necessarily actually having the lower bits duplicated in the host side register. This removes inefficient INS instructions from sequential single
float instructions.
In particular a very heavy single heavy block in Animal Crossing went from 712 instructions down to 520 instructions(~37% less instructions!)
If we are flushing multiple sequential guest GPRs then we can store two in a single STP instruction.
Ikaruga does this quite a bit in their blocks where they do an lmw at the very end and then we have to flush them all.
Typically cuts 16 STR instructions down to 8 STP instructions there.
This instruction is fairly heavily used by Ikaruga to load a bunch of registers from the stack.
In particular at the start of the second stage is a block that takes up ~20% CPU time that includes a usage of lmw to load half of the guest
registers.
Basic thing optimized here is changing from a single 32bit LDR to potentially a single 128bit LDR.
a single 32bit LDR is fairly slow, so we can optimize a few ways.
If we have four or more registers to load, do a 64bit LDP in to two host registers, byteswap, and then move the high 32bits of the host registers in
to the correct mapped guest register locations.
If we have two registers to load then do a 32bit LDP which will load two guest registers in a single instruction.
and then if we have only one register left to load, load it as before.
This saves quite a bit of cycles since the Cortex-A57 and A72's LDR instruction takes a few cycles.
Each 32bit LDR takes 4 cycles latency, plus 1 cycle for post-index(which typically happens in parallel.
Both the 32bit and 64bit LDP take the same amount of latency.
So we are improving latencies and reducing code bloat here.
This is a bug that crops if BindToRegister() is called multiple times in a row without a R() function call between them.
How to reproduce the bug:
1) Have a completely filled cache with no host register remaining
2) Call BindToRegister() with different guest registers
3) Don't call R() between the BindToRegister() calls.
This issue typically wouldn't be seen for a couple of reasons. Typically we have /plenty/ of registers in the cache, and in most cases we only call
BindToRegister() once per instruction. In the off chance that it is called multiple times, it wouldn't update the last used counts and would flush the
same register as the previous call to it.
