LLVM ERROR: Buffer too small for code

[James_R_]

When using the OpenCL compiler to compile a .cl file with a large number of kernels on a machine with both a Xeon CPU and Xeon Phi, the following message is dumped:
LLVM ERROR: Buffer too small for code

Also, the kernel doesn't finish compiling.  288 kernels may be a lot, but modern compilers should be able to handle this.

It does compile on a machine without a Xeon Phi.  It also compiles for Phi if I break the file into a several files with fewer kernels.

Can anyone from Intel repeat this bug?

[Frances_R_Intel]

The OpenCL folks are going to look at it and see if they can reproduce the problem. If they need any more information, we will get back to you. For now, please use your work around.

[Yuri_K_Intel]

Hi James, Is there any chance that you could share a reproducer (a kernel file and build options if any)? Otherwise, it will take some time for us trying to make up a one and there is no guarantee that the issue you're observing will actually be reproduced. Thanks, Yuri

[Yuri_K_Intel]

An update regarding the issue. We are observing the same error in one of our tests which also contains a file with a large number of kernels, so now we have a ticket and a reproducer for analysis. Thanks, Yuri

[James_R_]

Great! When can we expect a fix?

[Sumedh_N_Intel]

The OpenCL team has been able to root cause the issue, however there is no ETA for a fix yet. I'll keep you updated.