Backward compatibility of LLVM IR - ll/bc files

Quite often I get to work on an old bug, where an old ll/bc file is attached as a testcase. These files, in most cases (if not all), need to be converted somehow to the latest format, for the trunk version to be able to parse it without an error.

So a few questions arise:

  1. Is there a standard way to convert an old ll/bc to the latest? If not, what is the common approach for these cases?

  2. If there is no such tool, should we have one? (I think we should, of course.)

  3. What about automatically let the ll and bc Readers do the conversion (internally)?

  4. What about versions? Is there an RFC (probably old) for ll/bc file versioning? (I know there is an optional record for version in the bc file format, but it is not really utilized.)

Cheers,
Ehud.

I thought the binary bitcode reader was backwards compatible already. That’s what is documented here https://llvm.org/docs/DeveloperPolicy.html#ir-backwards-compatibility

I admit I didn’t know about that, but that is because I am handling a lot of old bugs, older than LLVM version 3.0.
Version 3.0 has been released in Dec 1st 2011, and there are still many bugs open from before that point.

The current BC Reader can’t parse files produced by LLVM version 2.9 and below (I’ve checked).
So, I wonder if there is anyone in favor, against (, or just doesn’t care) for supporting those versions?

Older releases are still available for download at releases.llvm.org; I believe the 3.0 release was supposed to be able to read 2.x bitcode, so you should be able to upgrade the bitcode with 3.0 tools and proceed from there. I think everything since 3.0 is still readable by current tools.

The project generally has not favored indefinite backward compatibility. There have not been any major bitcode format changes since then, so there has been little motivation to declare another compatibility break. Being able to read 9-year-old bitcode does not seem like a valuable feature, given the above workaround.

Let me say, though, that I fully support efforts to review and address old bugs. The LLVM project as a whole fixes (or otherwise closes) only about 2/3 of reported bugs. Improving that statistic would be very nice.

–paulr

Yes this is the process: use LLVM 3.0 to upgrade older bitcode to 3.0 format, and then 3.1 and above can load .bc files from 3.0.
The textual .ll files don’t have any support across version though, only the .bc.

Older releases are still available for download at releases.llvm.org; I believe the 3.0 release was supposed to be able to read 2.x bitcode, so you should be able to upgrade the bitcode with 3.0 tools and proceed from there. I think everything since 3.0 is still readable by current tools.

The project generally has not favored indefinite backward compatibility. There have not been any major bitcode format changes since then, so there has been little motivation to declare another compatibility break. Being able to read 9-year-old bitcode does not seem like a valuable feature, given the above workaround.

Let me say, though, that I fully support efforts to review and address old bugs. The LLVM project as a whole fixes (or otherwise closes) only about 2/3 of reported bugs. Improving that statistic would be very nice.

Agreed - though not sure what the chance of bugs that old still being relevant, given how much LLVM changes over time. If most of those older bugs are no longer relevant, it might be more useful to mass-close them and spend the effort that might’ve gone into reproducing/updating them into looking at more recent bugs. But certainly “dealer’s choice” - if looking at older bugs takes someone’s fancy especially, that’s cool :slight_smile:

  • Dave

On a lark, I skimmed the headlines of the 500 oldest open bugs. Sure a bunch of them are likely to be irrelevant now (I see some reported against dragonegg for example), but not all of them—some are still reasonable feature requests for example. I’d argue against a mass closure purely on the basis of age.

–paulr

On a lark, I skimmed the headlines of the 500 oldest open bugs. Sure a bunch of them are likely to be irrelevant now (I see some reported against dragonegg for example), but not all of them—some are still reasonable feature requests for example. I’d argue against a mass closure purely on the basis of age.
--paulr

This may be where a bot auto closing inactive issues might be useful.
The bot can ping the reporter and commenter and close the bug if there
is no response for a specified number of days.

From: Fangrui Song <maskray@google.com>
Sent: Tuesday, April 28, 2020 5:00 PM
To: Robinson, Paul <paul.robinson@sony.com>
Cc: David Blaikie <dblaikie@gmail.com>; llvm-dev <llvm-dev@lists.llvm.org>
Subject: Re: [llvm-dev] Backward compatibility of LLVM IR - ll/bc files

>On a lark, I skimmed the headlines of the 500 oldest open bugs. Sure a
bunch of them are likely to be irrelevant now (I see some reported against
dragonegg for example), but not all of them—some are still reasonable
feature requests for example. I’d argue against a mass closure purely on
the basis of age.
>--paulr

This may be where a bot auto closing inactive issues might be useful.
The bot can ping the reporter and commenter and close the bug if there
is no response for a specified number of days.

If there are any proposals to make in this direction, they should have
their own llvm-dev thread so the discussion is not buried in some
otherwise largely uninteresting-to-many-people thread.
--paulr