Well, there is still hope...
I think you should ask your customers in some way. And please provide my feeback to (the group of) people that will make the decision.
As for BFE, and more generaly, if the patterns you optimize are only the easy ones with constants.... then, sorry, this is useless for me (and probably not only me).
Lot's of project try to generate specialized opencl code. Instead of using lot's of layers of (useless) abstraction and code eg.: shader lang - > shader ir -> opencl codegen -> opencl compiler-> llvm (amdIL)-> hw IL-> hw why not simply cut these layers out?
Why not just release the llvm amdil backend specification with some kind of inline IL support with a callable llvm library backend ? Every high performance / heterogen computing project try to develop their own compiler and runtime nowdays - they will love it. Domain specific language developers (just check the shader language palette) will also love it. And suddenly we can port (almost) every language that support llvm to the gpu..
Please let us know what and how we can help to you to finish your work.I have some experience with llvm.
A direct llvm backend (with inline IL) for the gpu will be the best of both word. This backend will start lot's of university research on heterogenous computing, domain specific languages and application accelerators.
Please keep in mind that the AMD GPU have lot more computation resource  and we would like to map these resources to predicatable(!) direct computation without lot's of complex abstraction and code generation layer.
That would be great.
At some point I looked if it was possible to generate my own ISA code, but it would have been too time consuming (at least for free), and I bumped into another issue which is the binary format. The .text section is easy to understand with available documentation, but there is a .rodata section the loader probably uses to load constants + some GPU parameters I couldn't find information about.
If AMD provides a llvm backend, fully packaged to generate a loadable binary file, I would be very happy indeed.
So, intrinsics in the 2.5 and the llvm backend just after Very good!