Groups | Search | Server Info | Login | Register
Groups > comp.compilers > #3696
| From | Derek <derek@shape-of-code.com> |
|---|---|
| Newsgroups | comp.compilers |
| Subject | Re: Paper: CrossTL: A Universal Programming Language Translator with Unified Intermediate Representation |
| Date | 2025-09-02 12:45 +0100 |
| Organization | Compilers Central |
| Message-ID | <25-09-002@comp.compilers> (permalink) |
| References | <25-09-001@comp.compilers> |
John, > Invent a universal intermediate language so you just have to write a front > end or back end per input or output language. It works somewhat when the > inputs and outputs are semantically close enough, viz the LLVM > intermediate language, but has always hit the wall when the languages are > too different. The purpose of a universal intermediate language is to reduce the cost of 1-to-1 translation from an n^2 problem to a linear problem in n. An alternative approach is to brute force it by training an LLM on lots of code. "In the WMT25 competition, the model achieved first place in 30 out of the 31 language categories it participated in." https://github.com/Tencent-Hunyuan/Hunyuan-MT [That model translates natural languages, like Ukranian to Mongolian. I don't see how that's relevant to translating computer languages. In the former case, pretty close is useful, in the latter, pretty close is a program that doesn't work. -John]
Back to comp.compilers | Previous | Next — Previous in thread | Find similar
Paper: CrossTL: A Universal Programming Language Translator with Unified Intermediate Representation John R Levine <johnl@taugh.com> - 2025-09-01 11:16 +0200 Re: Paper: CrossTL: A Universal Programming Language Translator with Unified Intermediate Representation Derek <derek@shape-of-code.com> - 2025-09-02 12:45 +0100
csiph-web