Thanks for catching this bug.  I'm now a bit surprised this code work correctly before.
I have a couple questions/comments below.  The most important is possibly missing "max_outstanding_request" line.  Other than that, this patch looks good to me and points out we have more items to fix.

src/mem/ruby/system/GPUCoalescer.py (Diff revision 2)

This patch points out that we really should not have an icache or IFETCHes as part of the GPUCoalescer.  It is designed to only handle data accesses from the GPU CUs.
You don't necessarily need to remove this line and make all the other necessary changes, but if you don't, we'll get someone from AMD to make that change.

1. Andreas Sandberg Feb. 27, 2017, 1:32 a.m. (Feb. 27, 2017, 1:32 a.m.)

   I had a quick look at GPUCoalescer.cc and it seems like the icache pointer is used in the hit callback. I could remove it and add an assert. However, since I don't know what's going on in this part of the code base, you're probably better off doing it yourselves.

2. Brad Beckmann March 2, 2017, 8:52 a.m. (March 2, 2017, 8:52 a.m.)

   Sure thing.  We'll take care of it.  Thanks again for posting this patch and doing the work you have.

src/mem/ruby/system/GPUCoalescer.py (Diff revision 2)

Aren't you missing the following line:
max_outstanding_requests = Param.Int(16,

       "max requests (incl. prefetches) outstanding")

Show all issues

1. Andreas Sandberg Feb. 27, 2017, 1:32 a.m. (Feb. 27, 2017, 1:32 a.m.)

   That line should already be included from the original version of the file.

2. Brad Beckmann March 2, 2017, 8:52 a.m. (March 2, 2017, 8:52 a.m.)

   Oops...sorry about that.  I didn't realize we were already redefining that variable.  That actually seems like another problem this patch is fixing.  Thanks!