Enabling -O2 breaks Clash.Magic.clashSimulation

[leonschoorl]

When running clash on modules compiled by GHC, like in a starter-project setup, enabling -O2 can break Clash.Magic.clashSimulation.
Specifically it's the option -fspec-constr, which is implied by -O2.

Steps to reproduce:

$ stack new my-clash-project clash-lang/simple

Change contents of src/Example/Project.hs to:

{-# OPTIONS_GHC -fspec-constr #-}
-- {-# OPTIONS_GHC -ddump-spec #-}
-- {-# OPTIONS_GHC -ddump-simpl -dverbose-core2core #-}
module Example.Project where

import Clash.Prelude

topEntity :: Int
topEntity
 | clashSimulation = 111
 | otherwise = 222

$ stack run clash -- Example.Project --vhdl

or

$ cabal run clash -- Example.Project --vhdl

Result:

  result <= to_signed(111,64);

We're living in a simulation 😱

This doesn't happen when running clash directly on the source file.

[leonschoorl]

clashSimulation = True
{-# OPAQUE clashSimulation #-}

When enabled SpecConstr turns:

topEntity = case clashSimulation of
 False -> b
 True -> a

into

topEntity = a

@christiaanb does this violate the contract of OPAQUE?

[christiaanb]

I guess it follows the letter of the contract, but not the spirit...

Note [OPAQUE pragma]
~~~~~~~~~~~~~~~~~~~~
Suppose a function `f` is marked {-# OPAQUE f #-}.  Then every call of `f`
should remain a call of `f` throughout optimisation; it should not be turned
into a call of a name-mangled variant of `f` (e.g by worker/wrapper).

Since clashSimulation and its content are completely elided in the final result, technically, we see no name-mangled version of clashSimulation. i.e. we wouldn't be surprised if something with an OPAQUE annotation got elided as part of dead code elimination. That being said, I think in this particular case, GHC is doing something to the OPAQUE annoted binder that is non-desirable.

[christiaanb]

You could try compiling Clash.Magic with {-# OPTIONS_GHC -O0 #-} to see whether that helps.

[martijnbastiaan]

Maybe

clashSimulation = unsafePerformIO (fromMaybe True <$> lookupEnv "XXX_ISSUE_2736")
{-# OPAQUE clashSimulation #-}

Should be a readMaybe in there too, but I think the general idea should be clear :)

[christiaanb]

Together with @leonschoorl we discoved that

import GHC.Magic (noinline)

clashSimulation = noinline True
{-# OPAQUE clashSimulation #-}

is a workaround/proper fix. At least it's not something where we need to wait for a new GHC release with a changed behavior with respect to OPAQUE.

[rowanG077]

Related perhaps: #2570 ?

[leonschoorl]

@rowanG077 I don't think so, here the problem is happening inside a transformation inside GHC, with #2570 it seems to be inside transformations in clash