Why compilers have failed and what we can do about it

Keynote of LCPC'10

一直很奇怪，这么多人在研究如何用编译器实现程序的自动并行化，如果真的可以，为什么现在貌似还是人工并行为主要的工作，而没有真正feasible & executable的模型或者什么的可以使用，如果不行，为什么没有人干脆证明它是个NP-hard问题，这样就可以暂时避免前赴后继的时间和精力耗费其上……

Summary:

This PDF, almost indicated by its title, tells the reasons why compilers have failed in the auto-parallelization.

For the past 25 years, the compiler have accomplishments in

-- Instruction-level parallelism(ILP)

-- Memory-hierarchy optimization(optimize (im)perfectly nested DO-loops + dense arrays with loop transformations & program abstractions, which has something to do with polyhedral methods)

-- Performance portability

- Java: byte-code interpretation + just-in-time(JIT) compilation

- FFTW, SPIRAL: codelets + empirical search

- ATLAS: parameterized program + empirical search

For Auto-parallelization, we only succeeded in the vectorization of some dense matrix programs, while the Databases(SQL programmers and DBMS implementers) and numerical linear algebra fields hit it.

Lessons learnt:

-- Compilers are good at lowering abstraction level of program but poor at raising it

-- Need to figure out levels and software contracts between levels