half valve
I'm using flamegraph to try and optimize my playstation emulator. Assuming I am reading this right, it looks like the right shift operators are taking up a pretty large chunk of my instruction decode function time. This is surprising to me since shifts should be pretty fast. Is there a faster way to grab a subset of bits other than shift and mask?
This is compiled with --release and
[profile.release]
debug = true
in my Cargo.toml
For example, Rt is implemented like so
impl InstructionArgs for u32 {
#[inline(always)]
fn rt(&self) -> u8 {
((self >> 16) & 0x1F) as u8
}
...
honest needleBOT
crisp delta
^ you might consider using this
hidden badge
Under high optimization levels, sometimes instructions can get filed under unexpected instructions.
Can you look at the assembly of the hot part in perf?
half valve
Can you look at the assembly of the hot part in `perf`?
It doesn't like like it generated anything particularly crazy. Just a mov and a shift
Although I might be missing something. This is the first time I have used perf directly
hidden badge
So yeah, it's not the shift taking up time, it's the move right before it.
Perhaps the input %edi stalls waiting for an earlier operation to complete?
(Alas, I'm no expert in processors. My main knowledge is that doing more is generally slower than doing less, so you should find places where you can do less.)