| 1 |
Date: Wed, 20 Jun 2001 12:32:22 -0500 |
| 2 |
From: Vikram Adve <[email protected]> |
| 3 |
To: Chris Lattner <[email protected]> |
| 4 |
Subject: .NET vs. our VM |
| 5 |
|
| 6 |
One significant difference between .NET CLR and our VM is that the CLR |
| 7 |
includes full information about classes and inheritance. In fact, I just |
| 8 |
sat through the paper on adding templates to .NET CLR, and the speaker |
| 9 |
indicated that the goal seems to be to do simple static compilation (very |
| 10 |
little lowering or optimization). Also, the templates implementation in CLR |
| 11 |
"relies on dynamic class loading and JIT compilation". |
| 12 |
|
| 13 |
This is an important difference because I think there are some significant |
| 14 |
advantages to have a much lower level VM layer, and do significant static |
| 15 |
analysis and optimization. |
| 16 |
|
| 17 |
I also talked to the lead guy for KAI's C++ compiler (Arch Robison) and he |
| 18 |
said that SGI and other commercial compilers have included options to export |
| 19 |
their *IR* next to the object code (i.e., .il files) and use them for |
| 20 |
link-time code generation. In fact, he said that the .o file was nearly |
| 21 |
empty and was entirely generated from the .il at link-time. But he agreed |
| 22 |
that this limited the link-time interprocedural optimization to modules |
| 23 |
compiled by the same compiler, whereas our approach allows us to link and |
| 24 |
optimize modules from multiple different compilers. (Also, of course, they |
| 25 |
don't do anything for runtime optimization). |
| 26 |
|
| 27 |
All issues to bring up in Related Work. |
| 28 |
|
| 29 |
--Vikram |
| 30 |
|