Re: New Book: Optimizing Compilers for Modern Architectures

Toon Moene,
> dmjones wrote:
>
>> > Optimizing Compilers for Modern Architectures: A Dependence-based
>>
>> The title is a bit misleading. It should really be called everything
>> there is to know about data dependency analysis.
>>
>
> Do you think it would be a good reference to explain shortcomings in
> existing compilers ?


I'm not sure any compiler book could be used for this purpose.


> Obviously, I'm thinking of g77. It would be *very* handy to have some
> sort of reference text to point to in explaining: "Indeed, g77 doesn't
> do this sort of optimisation, <work-is-underway> | <here-is-why> |
><if-you-know-how-to-do-this-please-contribute>".


Using some of the optimizations discussed in the book would require
knowledge of specific processor/architecture details. For instance,
Chapter 9 discusses cache management. The size of cache is a
parameter in various equations used to unroll/interchange/tile loops.
If you are looking to generate truely optimal code then the cache size
needs to be a compile time option.


Is there sufficient demand for this level of developer involvement in
code generation? I know the people who really want performance are
often willing to pay lots of money, but it is a niche market. My
experience with companies who have Fortran source, these days, is that
they keep it because it does a job and converting to another language
is not cost effective.


I looked at the algorithms in this book and thought how much simpler
life would be, if C arrays were treated as arrays and it was not
possible to take their address.

Return to the comp.compilers page.
Search the comp.compilers archives again.