christiaanb ( Christiaan Baaij )


data Foo = F0 | F1 | F2 deriving (Generic, BitPack)

f :: Foo -> Unsigned 2
f F0 = 0
f F1 = 1
f F2 = 2

g :: Foo -> Unsigned 2
g F2 = 2
g F1 = 1
g F0 = 0

f' :: BitVector 2 -> Unsigned 2
f' = f . pack

g' :: BitVector 2 -> Unsigned 2
g' = g . pack

In simulation both behave identically and for 0b11 return 2
In HDL f' behaves like simulation
In HDL g' returns 0 for 0b11
IMHO the principle of least surprise would indicate that returned value should be errorX. This allows synthesis to produce more efficient implementation and matches behaviour of partial function in Haskell returning ⊥.

Catch-all clause

I'd note that sometimes the simulation-like behaviour is intended:


data MMIOReg = Reg0 | Reg1 | ... RegN | UnknownReg deriving (Generic)

{-# ANN module (DataReprAnn
                  $(liftQ [t|MMIO|])
                  32
                  [ ConstrRepr 'Reg0 0xFFF 0x000 []
                  , ConstrRepr 'Reg1 0xFFF 0x004 []
                  ...
                  , ConstrRepr 'RegN 0xFFF 0x??? []
                  , ConstrRepr 'UnknownReg 0x000 0xFFC []
                  ]) #-}

-- This works as intended
readMMIO :: MMIOState -> MMIOReg -> Unsigned 8
readMMIO st Reg0 = ...
readMMIO st Reg1 = ...
readMMIO st ... = ...
readMMIO _ UnknownReg = 0 -- For unknown values return 0

-- But this always return 0 in HDL (but not simulation)
readMMIO' :: MMIOState -> MMIOReg -> Unsigned 8
readMMIO' _ UnknownReg = 0 -- For unknown values return 0
readMMIO' st Reg0 = ...
readMMIO' st Reg1 = ...
readMMIO' st ... = ...

Proposed behaviour

Since both behaviors seems to have use-cases probably it would be best:

Always include errorX as last case (user cannot do it without having dead code warning from GHC)
On pattern match don't check for bit representation only but check for all previous states. In case of readMMIO' _ UnknownReg = 0 it would lower to readMMIO' _ r | not (reg0Pattern r || reg1Pattern r || ...) = 0; | otherwise = ...

created time in 3 hours

push eventclash-lang/stack-templates

Martijn Bastiaan

commit sha d73d13470ea88775782fc2235073f17c64bbec8f

Add simple-nix

view details

push time in 4 hours

create barnchclash-lang/stack-templates

branch : simple-nix

created branch time in 4 hours

issue commentclash-lang/clash-compiler

Use singletons instead of own SNat

singletons already depends on template-haskell, so perhaps the maintainers would accept a PR?

uzytkownik

comment created time in 8 hours

issue commentclash-lang/clash-compiler

romFile expects incorrect size of address bits

Has this been given up on? Since the current romFile is unusable, I don't think there's any backwards compatibility to be maintained.

GuzTech

comment created time in 11 hours

issue commentclash-lang/clash-compiler

Use singletons instead of own SNat

@gergoerdi Yes. The definitions are almost exactly the same:

Singletons:

data SNat (n :: Nat) = KnownNat n => SNat
type instance Sing = SNat

Clash:

data SNat (n :: Nat) where
  SNat :: KnownNat n => SNat n

uzytkownik

comment created time in 12 hours

issue commentclash-lang/clash-compiler

Use singletons instead of own SNat

Can you extract a KnownNat witness from a singletons-style Nat singleton? I would expect that to be the biggest hurdle.

uzytkownik

comment created time in 13 hours

issue openedclash-lang/clash-compiler

Use singletons instead of own SNat

singletons library is already in dependencies of clash. However clash uses own SNat implementation. That means that code using singletons needs to convert back and forth between them.

created time in 20 hours

push eventclash-lang/clash-starters

Clash DevOps

commit sha 1b2054a71711e98de658a783e7697a39150ec39c

Automated push from clash-lang/stack-templates

view details

push time in a day

push eventclash-lang/clash-compiler

Alex McKenna

commit sha 075085754d84b732b75ca9a4b4efeed192a3e6b9

WIP - Implement primops in the partial evaluator

view details