Targets/AMDGPU: Don't set fp32-denormals feature for r600

[jvesely]

Summary

This feature was removed in r335942

[arsenm]

According to cayman manual, these registers do exist so we should probably just make the feature accepted on r600 as well

[jvesely]

In D49650#1175438, @arsenm wrote:

According to cayman manual, these registers do exist so we should probably just make the feature accepted on r600 as well

sure, that's the way it was before r335942. I assumed the removal was intentional.

[arsenm]

In D49650#1175461, @jvesely wrote:

In D49650#1175438, @arsenm wrote:

According to cayman manual, these registers do exist so we should probably just make the feature accepted on r600 as well

sure, that's the way it was before r335942. I assumed the removal was intentional.

Probably accidental because nothing in r600 was actually using it

[jvesely]

In D49650#1176336, @arsenm wrote:

In D49650#1175461, @jvesely wrote:

In D49650#1175438, @arsenm wrote:

According to cayman manual, these registers do exist so we should probably just make the feature accepted on r600 as well

sure, that's the way it was before r335942. I assumed the removal was intentional.

Probably accidental because nothing in r600 was actually using it

given the number of warnings it outputs, I find that unlikely.
@tstellar what was your intention? It's not like someone is going to work on EG/CM denormals any time soon.

I don't mind either way. I just want to avoid another round of bikeshedding.

[jvesely]

In D49650#1177323, @jvesely wrote:

In D49650#1176336, @arsenm wrote:

In D49650#1175461, @jvesely wrote:

In D49650#1175438, @arsenm wrote:

According to cayman manual, these registers do exist so we should probably just make the feature accepted on r600 as well

sure, that's the way it was before r335942. I assumed the removal was intentional.

Probably accidental because nothing in r600 was actually using it

given the number of warnings it outputs, I find that unlikely.
@tstellar what was your intention? It's not like someone is going to work on EG/CM denormals any time soon.

I don't mind either way. I just want to avoid another round of bikeshedding.

D49934

lib/Basic/Targets/AMDGPU.cpp

     if (I == "+fp64-fp16-denormals" || I == "-fp64-fp16-denormals")
       hasFP64Denormals = true;
   }
-  if (!hasFP32Denormals)
+  if (!hasFP32Denormals && isAMDGCN(getTriple()))
     TargetOpts.Features.push_back(
         (Twine(CGOptsGPU.HasFastFMAF && !CGOpts.FlushDenorm
                    ? '+'