pulley: Implement iadd_pairwise

[eagr]

part of #9783

[github-actions]

Subscribe to Label Action

cc @fitzgen

This issue or pull request has been labeled: "cranelift", "pulley"

Thus the following users have been cc'd because of the following labels:

• fitzgen: pulley

To subscribe or unsubscribe from this label, edit the .github/subscribe-to-label.json configuration file.

Learn more.

[alexcrichton]

For these instructions we'll definitely want to avoid an intermediate allocation of a Vec. It might be reasonable to just use manual indices here and/or "macro expanding" some code. I don't think the standard library helpers for all this are necessarily the best way to go.

[Xuanwo]

We only have limited numbers of items to operate, maybe we can just write them out like:

fn vaddpairwisei16x8_s(&mut self, operands: BinaryOperands<VReg>) -> ControlFlow<Done> {
    let a = self.state[operands.src1].get_i16x8();
    let b = self.state[operands.src2].get_i16x8();
    
    let mut result = [0i16; 8];
    
    // Process first 4 elements from a
    result[0] = a[0].wrapping_add(a[1]);
    result[1] = a[2].wrapping_add(a[3]);
    result[2] = a[4].wrapping_add(a[5]);
    result[3] = a[6].wrapping_add(a[7]);
    
    // Process first 4 elements from b 
    result[4] = b[0].wrapping_add(b[1]);
    result[5] = b[2].wrapping_add(b[3]);
    result[6] = b[4].wrapping_add(b[5]);
    result[7] = b[6].wrapping_add(b[7]);

    self.state[operands.dst].set_i16x8(result);
    ControlFlow::Continue(())
}

This also has a compiler to perform better analysis, I suppose.

[eagr]

Guess I need a mind shift for this :)