In D24930#665944, @jvesely wrote:

In D24930#665738, @tstellarAMD wrote:

In D24930#665537, @jvesely wrote:

Why is this necessary? is the explicit mention of "merge ... sublicense and/or sell" the dealbraker?

The license is not an approved license for LLVM projects. If we want to add it back, we need to ask for an exception for this license from the LLVM foundation.

Is there a description how to do that? has the process been started?
Is there an alternative implementation with compatible license?

Feb 3 2017, 11:53 AM

• tstellarAMD added inline comments to D29473: [AMDGPU] Unroll preferences improvements.

Feb 3 2017, 8:33 AM

• tstellarAMD added inline comments to D29473: [AMDGPU] Unroll preferences improvements.

Feb 3 2017, 6:39 AM

• tstellarAMD added a comment to D24930: libclc: Remove code that is not licensed under either the UIUC or MIT license.

In D24930#665537, @jvesely wrote:

Why is this necessary? is the explicit mention of "merge ... sublicense and/or sell" the dealbraker?

Feb 3 2017, 6:07 AM

Feb 2 2017

• tstellarAMD added a comment to D20116: Add speculatable function attribute.

In D20116#664236, @mehdi_amini wrote:
LGTM. But let's wait a little to give a change to @sanjoy or @chandlerc to comment if they feel the need.

I repeat here one of the important earlier comment from @tstellarAMD , since it is not in the description and easy to miss:
I added two new Intrinsic attributes IntrNoSideEffects and IntrHasSideEffects,
which make it possible to specify all the possible memory interaction / side effect
combinations. With these properties in place, it should be possible in the future
to drop the 'no side effect' portion of the intrinsic memory properties once targets
have been updated to use these new properties.

Feb 2 2017, 12:19 PM

• tstellarAMD updated the diff for D24929: libclc: Remove copyright headers from top of files.

Add some build fixes from awatry.

Feb 2 2017, 9:53 AM

• tstellarAMD updated the diff for D24930: libclc: Remove code that is not licensed under either the UIUC or MIT license.

Also remove tgamma which calls lgamma.

Feb 2 2017, 9:51 AM

• tstellarAMD updated the diff for D29452: AMDGPU/SI: DAGMutation for removing deps between CSEs.

Correctly remove predecessors and clean up some of the loops.

Feb 2 2017, 9:15 AM

• tstellarAMD created D29452: AMDGPU/SI: DAGMutation for removing deps between CSEs.

Feb 2 2017, 7:07 AM

Feb 1 2017

• tstellarAMD accepted D20552: AMDGPU: Allow clustering flat memory operations.

Feb 1 2017, 10:59 AM

• tstellarAMD added a comment to D28993: AMDGPU: Try to select SMEM opcodes for llvm.amdgcn.buffer.load.

Another possible issue is that SMEM instructions ignore bits of the resource descriptor. So you would need some way to tell the compiler that it wouldn't be ignoring some relevant resource bits by selecting to SMEM.

Feb 1 2017, 10:16 AM

• tstellarAMD added a comment to D27586: AMDGPU/SI: Add llvm.amdgcn.s.buffer.load intrinsic.

In D27586#663090, @mareko wrote:

Would you please describe the purpose of this patch? It's not obvious why it's useful.

Feb 1 2017, 9:59 AM

• tstellarAMD updated the diff for D20116: Add speculatable function attribute.

Update definition in LangReg.

Feb 1 2017, 9:35 AM

• tstellarAMD added a comment to D27586: AMDGPU/SI: Add llvm.amdgcn.s.buffer.load intrinsic.

In D27586#662991, @nhaehnle wrote:

I haven't looked in too much detail yet. I assume getelementptr doesn't work with these pointers, so it would be good to have a negative test which ensures that GEP use fails.

Feb 1 2017, 3:56 AM

Jan 31 2017

• tstellarAMD added a comment to D28993: AMDGPU: Try to select SMEM opcodes for llvm.amdgcn.buffer.load.

What if something else has written to the buffer in the same shader? That would make using smem instructions illegal.

Jan 31 2017, 5:37 PM

• tstellarAMD abandoned D6820: Add LLVMSetTargetMachineMCOption() to the C API.

Jan 31 2017, 5:36 PM

• tstellarAMD added a comment to D27586: AMDGPU/SI: Add llvm.amdgcn.s.buffer.load intrinsic.

In D27586#662432, @tstellarAMD wrote:

In D27586#662394, @mareko wrote:

How is this different from using amdgcn.buffer.load if D28993 lands (which is not certain)?

I don't think it's legal to select amdgcn.buffer.load to SMRD unless you can prove that it is uniform. llvm.amdgcn.s.buffer.load is known to always be uniform.