NIR Still Being Discussed For Mesa, LLVM Gets Brought Up Again

Last week NIR was announced as a new intermediate representation for Mesa. Discussion around this new but experimental IR continues to happen among upstream developers while Mesa developers are back to discussing LLVM too.

Mesa developers appear interested in NIR but it's not clear yet that anyone is committed to finishing up and hooking up to the drivers this new intermediate representation that's lower-level than GLSL IR that supports SSA and is typeless. Those wishing to learn more of the gritty details being discussed can browse through the lengthy mailing list thread.

Along with NIR, Mesa developers are back to discussing LLVM and its IR possibilities for Mesa. While LLVM is used extensively by Radeon (R600g/RadeonSI) and LLVMpipe drivers most notably, Intel has been the main Mesa contributor against LLVM. At multiple stages of Intel's Linux graphics driver history with Mesa had been plans to develop an Intel LLVM back-end and other LLVM investigations for Intel's GPU driver, but in the end they've become opposed to LLVM.

Their opposition to using LLVM in their graphics driver comes down to the significant work in rewriting their compiler back-end and the significant amount of upfront work they must invest in this back-end before seeing any return on investment with LLVM's optimization potential, etc. The Intel driver has also seen a bumpy road in the past when reworking large parts of their Linux graphics driver stack and are looking to avoid such fallout going forward.

Besides Intel not being interested in the upfront challenge and the headaches involved in learning the ins and outs of LLVM, other common issues with LLVM Mesa usage come down to LLVM's lack of a stable API (it generally breaks release-to-release), the different LLVM release schedule, and packaging challenges for distribution vendors, among other potential issues. It's been brought up again in this NIR mailing list thread about potentially bundling the LLVM source tree within Mesa, but that too entails a different set of headaches.

The latest in the NIR thread proposes using LLVM's C API to write a translator from Mesa's GLSL IR into LLVM IR, utilize LLVM's optimization abilities, and then give the driver back-ends the option of using the LLVM IR. This approach would at least allow once and for all the Mesa developers to see in close detail how useful LLVM's optimizers are in practice for Mesa. However, no developers have yet committed themselves to this work.

VMware isn't interested in contributing resources to this effort but Jose Fonseca of VMware noted in the NIR discussion, "even though it seems to me that Mesa will unavoidably eventually gravitate towards more wide-spread use of LLVM, I'm unable to help provide more critical mass myself. Which is why I don't oppose NIR. Even though it seems a less direct path than overcome the hurdles of using LLVM now, NIR's SSA-nature seems a step forward nevertheless. And of course, who does the job gets to decide how it's done."

Also worth bringing up again is the interesting work done by LunarG with LunarGLASS and the Valve-funded work on Glassy Mesa for incorporating LLVM into Mesa for offering increased driver performance.

In the end the usage of LLVM inside Mesa is most likely to grow and the existing challenges overcome, it's just a matter of time.