| 1 |
Date: Sat, 18 Nov 2000 09:19:35 -0600 (CST) |
| 2 |
From: Vikram Adve <[email protected]> |
| 3 |
To: Chris Lattner <[email protected]> |
| 4 |
Subject: a few thoughts |
| 5 |
|
| 6 |
I've been mulling over the virtual machine problem and I had some |
| 7 |
thoughts about some things for us to think about discuss: |
| 8 |
|
| 9 |
1. We need to be clear on our goals for the VM. Do we want to emphasize |
| 10 |
portability and safety like the Java VM? Or shall we focus on the |
| 11 |
architecture interface first (i.e., consider the code generation and |
| 12 |
processor issues), since the architecture interface question is also |
| 13 |
important for portable Java-type VMs? |
| 14 |
|
| 15 |
This is important because the audiences for these two goals are very |
| 16 |
different. Architects and many compiler people care much more about |
| 17 |
the second question. The Java compiler and OS community care much more |
| 18 |
about the first one. |
| 19 |
|
| 20 |
Also, while the architecture interface question is important for |
| 21 |
Java-type VMs, the design constraints are very different. |
| 22 |
|
| 23 |
|
| 24 |
2. Design issues to consider (an initial list that we should continue |
| 25 |
to modify). Note that I'm not trying to suggest actual solutions here, |
| 26 |
but just various directions we can pursue: |
| 27 |
|
| 28 |
a. A single-assignment VM, which we've both already been thinking about. |
| 29 |
|
| 30 |
b. A strongly-typed VM. One question is do we need the types to be |
| 31 |
explicitly declared or should they be inferred by the dynamic compiler? |
| 32 |
|
| 33 |
c. How do we get more high-level information into the VM while keeping |
| 34 |
to a low-level VM design? |
| 35 |
|
| 36 |
o Explicit array references as operands? An alternative is |
| 37 |
to have just an array type, and let the index computations be |
| 38 |
separate 3-operand instructions. |
| 39 |
|
| 40 |
o Explicit instructions to handle aliasing, e.g.s: |
| 41 |
-- an instruction to say "I speculate that these two values are not |
| 42 |
aliased, but check at runtime", like speculative execution in |
| 43 |
EPIC? |
| 44 |
-- or an instruction to check whether two values are aliased and |
| 45 |
execute different code depending on the answer, somewhat like |
| 46 |
predicated code in EPIC |
| 47 |
|
| 48 |
o (This one is a difficult but powerful idea.) |
| 49 |
A "thread-id" field on every instruction that allows the static |
| 50 |
compiler to generate a set of parallel threads, and then have |
| 51 |
the runtime compiler and hardware do what they please with it. |
| 52 |
This has very powerful uses, but thread-id on every instruction |
| 53 |
is expensive in terms of instruction size and code size. |
| 54 |
We would need to compactly encode it somehow. |
| 55 |
|
| 56 |
Also, this will require some reading on at least two other |
| 57 |
projects: |
| 58 |
-- Multiscalar architecture from Wisconsin |
| 59 |
-- Simultaneous multithreading architecture from Washington |
| 60 |
|
| 61 |
o Or forget all this and stick to a traditional instruction set? |
| 62 |
|
| 63 |
|
| 64 |
BTW, on an unrelated note, after the meeting yesterday, I did remember |
| 65 |
that you had suggested doing instruction scheduling on SSA form instead |
| 66 |
of a dependence DAG earlier in the semester. When we talked about |
| 67 |
it yesterday, I didn't remember where the idea had come from but I |
| 68 |
remembered later. Just giving credit where its due... |
| 69 |
|
| 70 |
Perhaps you can save the above as a file under RCS so you and I can |
| 71 |
continue to expand on this. |
| 72 |
|
| 73 |
--Vikram |
| 74 |
|