LLVM 4.0 Causes Slow Performance For RadeonSI?

[dungeon]

Workstations goes with 4 times per year driver updates, APUs are in similar scheme... so only gamers with cards get planty of i can't wait alpha/beta

[dungeon]

So it is similar to Linux with mese with 3-4 stable drivers per year... and a lot of hotfixes in between, for those who want them of course as that is optional there and here

[dungeon]

Funny, isn't it. It reflects how important the latest dev versions are to people. It almost feels like nobody cares about or tests the stable versions anymore. Poor distributions.

He, he, i don't think so . Gaming people wanna latest dev version because stable is not really stable so lets try new because "it might fix something" approch... just imagine TF2 does not lockup GPUs with stable, most people would not have a need to try someting new, but currently they hope newset will just fix some of those bugs

And worse thing is when they found issue is not fixed even there than "OK, once i am here i will stay here and wait" Basically, it is Waiting for GoDoT game and big bugs free drivers all around

[puleglot]

Even Mesa 12 contains "#if HAVE_LLVM >= 0x0309" conditional compiler directives, so that statement from him didn't make sense at all. Mesa and LLVM goes in lockstep.

mesa-12.0 has been branched when llvm-3.9 was in development. Therefore, it contains "#if HAVE_LLVM >= 0x0309" in various places. llvm-3.9 has been branched later and at this point it contained changes that made it slightly incompatible with mesa-12.0. So what's wrong with that statement?

[puleglot]

Funny, isn't it. It reflects how important the latest dev versions are to people. It almost feels like nobody cares about or tests the stable versions anymore. Poor distributions.

I always use latest stable versions of mesa and llvm. But I'm on gentoo.

[zboszor]

mesa-12.0 has been branched when llvm-3.9 was in development. Therefore, it contains "#if HAVE_LLVM >= 0x0309" in various places. llvm-3.9 has been branched later and at this point it contained changes that made it slightly incompatible with mesa-12.0. So what's wrong with that statement?

The "can't support unreleased dependency" statement means that you don't develop for it and you don't test with it.
When said dependency has released a new version then you start developing for it.

But this is not what's happening in Mesa and therefore that statement is hypocrisy.

[MrCooper]

The "can't support unreleased dependency" statement means that you don't develop for it and you don't test with it.
When said dependency has released a new version then you start developing for it.

But this is not what's happening in Mesa and therefore that statement is hypocrisy.

The fundamental issue is that LLVM doesn't provide API/ABI compatibility between its releases. The API/ABI can change in arbitrary ways at any time, and does fairly frequently. As long as that's the case, my statement will be true (in general, though there could be exceptions if releases of both projects are branched off the main development branch around the same time, but the Mesa release ends up earlier).

The Mesa Git master branch is constantly being adapted to API/ABI changes in LLVM SVN trunk. This is required for catching regressions with development snapshots of both projects as early as possible. But once a Mesa release branches off master, it stops getting those updates. The longer after that an LLVM release branches off SVN trunk, the smaller the chance that the Mesa release branched earlier will even compile with it, let alone work correctly.

And before someone says "just backport the fixes for the new LLVM release to the Mesa release": It's not always easy to identify all such changes, see e.g. the last comment on the bug report you referenced: https://bugs.freedesktop.org/show_bug.cgi?id=97542#c12