Subject: Re: Why do we put GMP allocations on GHC heaps?
Date: Wednesday 23rd October 2013 16:49:00 UTC (over 2 years ago)
First, I took some time a year or two ago to give a go at writing an integer-openssl implementation. I don't have the code anymore (it was more of a fun exercise,) sorry. :( However, the real problem with the whole integer situation IMO, is that the choice of integer implementation is completely non-modular, everything else aside. In a theoretically ideal world, it would be a link-time option, e.g. $ ghc foo.hs $ ghc -threaded foo.hs will relink `foo` with the -threaded runtime. Similarly we would like to say: $ ghc foo.hs $ ghc -integer-type=gmp foo.hs $ ghc -integer-type=openssl foo.hs $ ghc -integer-type=simple foo.hs and so on. This is besides all of the talk of "we need to tell the GC that a root is here, so it doesn't free things, etc." as in the case with MPFR. The current scheme is non-modular because it depends on a package-level interface which the compiler cannot hide. Furthermore, it depends on a specific build of GHC to choose, and various people end up needing this but do not want to recompile (for the standard - valid - typical deployment reasons: you want a stable version of your compiler, not a home-grown one.) Unfortunately the current situation with a package-level interface has various ramifications making this theory quite difficult in practice. A) Because it's an exposed-package, people actually do, sometimes, depend on it or the API in various ways. This means swapping implementations becomes more difficult (OTOH, such low-level-ness is a rare occurrence for most users, so perhaps this wouldn't hurt too many people.) B) The bigger problem related to A) is that because it's an exposed-package, GHC will inline the hell out of unfoldings for integer-gmp primitives at call sites. That means a compiled module may get an inlined copy of some integer-gmp primitive for example (a library could use Integer internally,) making re-linking later basically impossible. C) Probably other things I can't think of off the top of my head. In particular I don't know how to do this and reconcile A and B without destroying the efficiency of the GMP-related primitives, or highly restricting the Integer implemenation API and specializing it further in the compiler to handle stuff like this. Or something. The idea is that we inline all the unfoldings to minimize the overhead at call sites, but I'm not sure the best way to do this at the linker level - in particular we might need to stub things out per-way, because relinking requires the same symbols both ways, and we want an interface that does not cost us much. It doesn't cost too much for the RTS because the bits to relink are fairly minimal and don't hurt the fast path very much, AFAIK. If half of the problem is actually integration issues, can I ask why we don't just pick a small C library and offer that up? It is highly likely we will *never* beat GMP unfortunately. Beating it in some specific cases might work, but it has been ruthlessly optimized for years by many skilled people, and I highly doubt we can beat it on average without a very, very substantial amount of work. It would be interesting work - but I'm not convinced we can't have an acceptable solution in the mean time (and by all means, should we get something faster, it can just replace integer-simple.) As an example, libtomcrypt is damn fast, BSD3 licensed, very small, and well-respected. Why don't we: 1) Take libtomcrypt, modify the symbol names to be private to GHC (e.g. prefix everything with `ghc_integer_`) - this means the symbol names will be internal to GHC, so this also doesn't affect relinks against other copies of libtomcrypt (many projects include their own copy.) It is so small it basically comes under our control completely. 2) We can then offer people a integer-tomcrypt which does not have compatibility issues, and does not cause as much pain as the integration of widely deployed libraries like GMP/OpenSSL, which are in various use by many Haskell packages already. This should give us the flexibility of integer-simple without compromising so much on the efficiency aspect. This is also a very half-baked idea, but just a suggestion. I'd love a more modular solution to Integer as well, as I know many other people would. On Wed, Oct 23, 2013 at 11:08 AM, Bryan O'Sullivan