Add blackbox for 'Clash.Sized.Vector.iterateI'

[martijnbastiaan]

Adds blackbox (and associated reduceNonRepPrim/evaluator rule) for Clash.Sized.Vector.iterateI. Works around #1240.

---

Name	Before (ms)	After (ms)	B/A
FIR	35.73	34.61	1.03
Reducer	771.6	732	1.05
Queens	16.89	16.30	1.03
BundleMapRepeat	21.70	16.88	1.28
PipelinesViaFolds	1083	208.1	5.20
AES	326.6	281.6	1.07
T1354B	∞	508

[alex-mckenna]

A couple of questions about things, but generally looks good. One question that didn't make sense being tagged in a particular area: does this work in Verilog / SystemVerilog? Some parts are explicitly VHDL instead of any backend, but other parts seem to be over every backend.

[martijnbastiaan]

One question that didn't make sense being tagged in a particular area: does this work in Verilog / SystemVerilog? Some parts are explicitly VHDL instead of any backend, but other parts seem to be over every backend.

Yes this works for all HDLs. The changes regarding the this are just an improvement on the DSL that was donated/upstreamed by Chris over at #1352. I haven't used any of those function for the iterateI blackbox though :)

[martijnbastiaan]

Thanks for the review @alex-mckenna! I've implemented your suggestions.

[martijnbastiaan]

Converting to draft PR. The following example makes Clash get stuck in an infinite loop:

module Test where

import Clash.Prelude

f :: forall n. KnownNat n => Index (n + 1) -> Int -> Int
f n = foldl (.) id lanes
 where
  lanes :: Vec (n + 1) (Int -> Int)
  lanes = map (\i -> if i < n then id else id) (iterateI succ 0) 

topEntity :: Int -> Int
topEntity = f (1 :: Index 5) 

[martijnbastiaan]

Testing on a private, bigger codebase revealed a 20% speedup @christiaanb :). I've applied the review suggestions, so I think this is good to go. (If merging, it's probably best to create a merge commit for this one, as there's quite a lot of info in this thread.)

[christiaanb]

Any reason not to backport to 1.2?

[martijnbastiaan]

It's a bit of work, most of which I've done in backport-workaround1240, but it needs updating to include the latest force pushes.

[christiaanb]

@martijnbastiaan Doesn't this duplicate (at least compiler) work for the vec term?

[martijnbastiaan]

Possibly / probably. How would you write it?

There should also be a "quick" path that just extract the element if it sees a data constructor.