Block Objects: Marshalling Them Explicitly

In the following, we use the standard Haskell 2010 FFI to explicitly marshal Haskell functions to C blocks and vice versa.

The qsort_b example using a foreign wrapper

The most straight-forward approach is to use the existing FFI support for turning Haskell functions into C function pointers by way of a foreign import wrapper declaration. We can then embed these C function pointers in block literals without the need for a explicit environment on the C side.

Block literals

Blocks are by default allocated on the stack in C and only promoted to the heap using an explicity Block_copy() function. In Haskell, we always allocate them on the heap, which is indicated by the pointer being placed in a block literals isa field:

foreign import ccall "& _NSConcreteGlobalBlock" nsConcreteGlobalBlock :: Ptr ()

A block literal with an empty environment has the following layout:

-- Layout of the block literal (64-bit runtime)
--
-- .quad	__NSConcreteGlobalBlock           # void *isa;
-- .long	1342177280                        # int  flags = 0x50000000;
-- .long	0                                 # int  reserved;
-- .quad	___block_invoke                   # void (*invoke)(void *, ...);
-- .quad	___block_descriptor               # struct Block_descriptor *descriptor;

long, quad :: Int
long = 4  -- long word = 32 bit
quad = 8  -- quad word = 64 bit

isaOffset, flagsOffset, invokeOffset, descriptorOffset, blockLiteralSize :: Int
isaOffset        = 0
flagsOffset      = isaOffset        + quad
invokeOffset     = flagsOffset      + long + long
descriptorOffset = invokeOffset     + quad
blockLiteralSize = descriptorOffset + quad

In Haskell, we represent block literals as opaque pointers:

newtype Block a = Block (Ptr (Block a))

When turning a Haskell function into a C function pointer to be included in a block literal as the invoke function, we need to take care to add a pointer to the block literal itself as a new first argument:

mkBlock :: ((Block f -> f) -> IO (FunPtr (Block f -> f))) -> f -> IO (Block f)
mkBlock mkWrapper f
  = do { fPtr     <- mkWrapper (const f)
       ; blockPtr <- mallocBytes blockLiteralSize
       ; poke (blockPtr `plusPtr` isaOffset)        nsConcreteGlobalBlock
       ; poke (blockPtr `plusPtr` flagsOffset)      (0x50000000 :: Word32)
       ; poke (blockPtr `plusPtr` invokeOffset)     fPtr
       ; poke (blockPtr `plusPtr` descriptorOffset) descriptorPtr
       ; return $ Block blockPtr
       }

The block descriptor is static, except for the signature that we omit for the moment.

-- Block descriptor structure shared between all blocks.
--
-- .quad	0                                 # unsigned long int reserved;
-- .quad	32                                # unsigned long int size = blockLiteralSize;
-- .quad	signature_str                     # const char *signature;
-- .quad	0                                 # <undocumented>

descriptorPtr :: Ptr ()
descriptorPtr
  = unsafePerformIO $ 
    do { descPtr <- mallocBytes (4 * quad)
       ; poke (descPtr `plusPtr` (0 * quad)) (0 :: Word64)
       ; poke (descPtr `plusPtr` (1 * quad)) blockLiteralSizeWord64
       ; poke (descPtr `plusPtr` (2 * quad)) nullPtr    -- gcc puts a NULL in; should be ok for now
       ; poke (descPtr `plusPtr` (3 * quad)) (0 :: Word64)
       ; return descPtr
       }
  where
    blockLiteralSizeWord64 :: Word64
    blockLiteralSizeWord64 = fromIntegral blockLiteralSize

TODO:

• Produce an example marshalling a block from C to Haskell
• Evaluate which parts of the code actually need to be generated and which parts could go into an extension of the FFI library (Foreign.C.Blocks) or into the RTS if it is C code.