Message dug out from 27/10/97, which nobody replied to... (Sorry
Sergey!)
[EMAIL PROTECTED] (S.D.Mechveliani) writes:
> When ghc-2.08-linux-i386..., (binaries by S.Panne)
>
> compiles the below module with -O, it in-lines both g and f,
> as this can be seen from M.hi
> But f is rather large to be in-lined, is not it?
>
> I tried to surpress it with {-# notINLINE f #-}.
> But ghc ignores it.
To suppress or encourage inlinings, the following flags may be useful:
-funfolding-use-threshold<n>
A metric governing the maximum size an expression can be
for it to be inlined. GHC < 3.02 set this to 3, we're
now setting it to 8.
-funfolding-con-discount<n>
A discount given to the cost of an inlining if the compiler
knows it can eliminate a 'case' expression by performing
the inlining. GHC < 3.02 set this to 1, we're now setting it
to 3.
-funfolding-keenness-factor<n>
All the discounts are multiplied by this factor. Default: 2.
-funfolding-creation-threshold<n>
A metric governing how big an expression can be for it to
be even *considered* for future inlinings. Default: 30.
-funfolding-interface-threshold<n>
The maximum size of a function definition for its unfolding
to appear in the interface file. (I think this is what
Sergey was after above). Default: 30.
GHC 3.02 will accept all of these flags, to use them with a previous
version just prefix the flag with '-optC'. GHC 3.02 should also also
have a 'NOINLINE' pragma if I get it working in time.
Cheers,
Simon
--
Simon Marlow [EMAIL PROTECTED]
University of Glasgow http://www.dcs.gla.ac.uk/~simonm/
finger for PGP public key
