3 ms·

Isn't this almost what LLVM is doing - creating a common instruction set for language designers to target and which compiles to a large amount of machines? Sry

by olodus 9y ago

Isn't this almost what LLVM is doing - creating a common instruction set for language designers to target and which compiles to a large amount of machines? Sry if it is stupid question. I'm interested in programming language design but are still very much learning the ropes. I might have misunderstood the real problem LLVM or this wants to solve.

thechao 9y ago
It's the same idea, yes. Although, the specification of UNCOL is... a little odd. It has a bunch of cells (referenced by address); instructions refer to a previous instruction, and an address for their arguments---so like a stack machine, but the second argument is always a pointer. Or ... some such thing. The one paper I read spends 80% of its effort defining the representation of text.
- msla 9y ago
  > The one paper I read spends 80% of its effort defining the representation of text. These days, an equivalent effort would incorporate Unicode and UTF-8 by reference, which together amount to much more verbiage, especially Unicode. Back then, text standards weren't quite so mature, and it's possible nothing existing at the time really solved their problem.
glangdale 9y ago
From what I recall, this is called out as a specific non-goal in the LLVM documentation. I don't remember the details but I think that the instruction set is meant to be an easy way to write passes that communicate with each other but is not designed to be a universal intermediate format.
abecedarius 9y ago
Web Assembly seems closer: there's less undefined behavior, as I understand it. (I've read much of the wasm spec but not LLVM docs.) https://github.com/WebAssembly/design/blob/master/Rationale.md https://github.com/WebAssembly/design/blob/master/Rationale....
pjmlp 9y ago
That has been done multiple times before LLVM. Most mainframe assembly is referred to as bytecode in old papers, because CPUs were micro-programmed, some of them you could already load such micro-program on boot. For example on this Alto demo they upload the micro-program to enable the Smalltalk environment. Notorious common instruction sets are P-Code, JVM, MSIL, TIMI.

by olodus 9y ago

thechao 9y ago

It's the same idea, yes. Although, the specification of UNCOL is... a little odd. It has a bunch of cells (referenced by address); instructions refer to a previous instruction, and an address for their arguments---so like a stack machine, but the second argument is always a pointer. Or ... some such thing. The one paper I read spends 80% of its effort defining the representation of text.

msla 9y ago
> The one paper I read spends 80% of its effort defining the representation of text. These days, an equivalent effort would incorporate Unicode and UTF-8 by reference, which together amount to much more verbiage, especially Unicode. Back then, text standards weren't quite so mature, and it's possible nothing existing at the time really solved their problem.

glangdale 9y ago

From what I recall, this is called out as a specific non-goal in the LLVM documentation. I don't remember the details but I think that the instruction set is meant to be an easy way to write passes that communicate with each other but is not designed to be a universal intermediate format.

abecedarius 9y ago

Web Assembly seems closer: there's less undefined behavior, as I understand it. (I've read much of the wasm spec but not LLVM docs.) https://github.com/WebAssembly/design/blob/master/Rationale.md https://github.com/WebAssembly/design/blob/master/Rationale....

pjmlp 9y ago

That has been done multiple times before LLVM. Most mainframe assembly is referred to as bytecode in old papers, because CPUs were micro-programmed, some of them you could already load such micro-program on boot. For example on this Alto demo they upload the micro-program to enable the Smalltalk environment. Notorious common instruction sets are P-Code, JVM, MSIL, TIMI.