|
| From: | Dan Sugalski |
| Subject: | Re: [Cardinal-dev] Compiler structure / AST format |
| Date: | Wed, 1 May 2002 15:07:44 -0400 |
At 8:14 AM +0300 4/26/02, Einar Karttunen wrote:
On 25.04 11:53, Dan Sugalski wrote:At 4:05 AM +0300 4/25/02, Einar Karttunen wrote: >The reason for the immediate is the optimiser pass and the >complexity of backends. >* backends need to only know about the intermediate code not the AST >* frontends and backends can be switched with no recoding >* some optimisations common to all backends are very unnatural at >the AST level I'll point out that some optimisations common to all backends are very natural to do at the AST level. Most compilers I know of have at least three different intermediate representations of the code for the optimiser because of the different characteristics of the different optimisations. Some work best with a high-level representation, some with a medium-level representation, and some with a low-level representation.Of course some optimisation is best done in AST level, and I think my message made it clear. Most compilers I know about have just two level if you don't count instruction scheduling as a third level.
It's worth counting instruction scheduling. (As I reply really tardily) It can make a significant difference, even more so when the instructions are as heavy-weight as you'll find in an interpreter. Still, they matter on hardware as well.
I wouldn't be surprised to see some cases where proper instruction scheduling on parrot, assuming the source languge will allow it, brings a win of 10-30% in execution time. Not hugely common, mind, but a definite possibility. (The smaller the L1 and L2 caches on a processor the more instruction scheduling will get you a performance win)
--
Dan
--------------------------------------"it's like this"-------------------
Dan Sugalski even samurai
address@hidden have teddy bears and even
teddy bears get drunk
| [Prev in Thread] | Current Thread | [Next in Thread] |