The optimizations he needs to make to the C code didn't seem clear to him when he was writing it, but they're very common optimizations for someone with more experience writing in C.
That he made the seemingly-natural optimization in Haskell while not affording C the same luxury is what's hurting his argument.
It is, of course, not the case that Haskell is faster than C with arbitrary effort expended tuning to the specific hardware - no one is claiming that.
import Data.Vector as V
import Data.Vector.Unboxed as V
My complaint is that the C code isn't given the same chance. He even calls out that reading data with getc is a known performance problem, but then does it anyway. Any book on learning C will point out that getc is slow for reading lots of data, and fscanf or fread should be used instead.
f* calls are buffered this is true. But depending on your workload this may be a blessing in disguise because if you're parsing a very long stream of unstructured data you'll have to re-implement most of that buffering logic in your own code to deal with the inevitable corner cases where your data overlaps the borders of the buffersize chosen. The obvious optimization here is to pre-allocate a buffer of the right size and read in the data in one go but for a stream of unknown and possibly infinite length (such as used in a filter like this) that is not possible.
So f* function have their place, even in optimized code, but should be used with care. If you're reading data with fixed size blocks (such as in this example) then yes, a simple 'read' call will likely be faster, because as you note the read call will place the data directly into a user accessible buffer without the need to go through more library calls to get at the data. (fread is a wrapper function around the read system call).
The pointer arithmetic he's using shouldn't need to be optimized, since the page sizes he's malloc'ing and pulling from memory are small enough that they should stay in the L1/L2 cache for the entire run(he's using 1k blocks of data, most processors use 4k pages). There's almost no optimization to be done there.
The biggest performance hit is actually the single character puts versus a block read or write.
Haskell is probably implemented to read a large block of the file(or perhaps the entire file) into memory, then parse after. That would be a minimal number of system fread calls over the entire run. Versus the C code, where 1 block's parse could be 1024 getc and putc calls.