AMDGPU: Remove AMDIL.round.nearest intrinsic

[nhaehnle]

The change makes sense, but we should note somewhere (mostly for distribution packagers) that this removes support for Mesa 11.0.x. Not sure where best to note this - definitely in the commit message, but probably also in some release notes file.

[arsenm]

In D16328#331386, @nhaehnle wrote:

The change makes sense, but we should note somewhere (mostly for distribution packagers) that this removes support for Mesa 11.0.x. Not sure where best to note this - definitely in the commit message, but probably also in some release notes file.

Is it expected that released versions of mesa will work with future llvm releases?

[arsenm]

I noted the version break in one of the earlier commit logs. D16368 adds a release note change

[michel.daenzer]

In D16328#331398, @arsenm wrote:

Is it expected that released versions of mesa will work with future llvm releases?

No, it's not. So documenting this wasn't really necessary, but thanks anyway. :)

[nhaehnle]

I don't think we publish "known working combinations" anywhere, and I do believe we should at least document it - thanks for that!

(And as an aside, I feel we should absolutely avoid changes to LLVM that break Mesa silently. This current change is okay because it isn't silent: using old Mesa will crash and burn with a rather explicit message as soon as the removed intrinsic is encountered. But it's theoretically conceivable that some more subtle LLVM change around ABIs could cause incorrect rendering without an explicit message about it, and that would be bad.)

[arsenm]

r258346

[michel.daenzer]

In D16328#332364, @nhaehnle wrote:

(And as an aside, I feel we should absolutely avoid changes to LLVM that break Mesa silently.

Of course.

This current change is okay because it isn't silent: using old Mesa will crash and burn with a rather explicit message as soon as the removed intrinsic is encountered. But it's theoretically conceivable that some more subtle LLVM change around ABIs could cause incorrect rendering without an explicit message about it, and that would be bad.)

Old Mesa generally doesn't build against new LLVM, which is why that's not a supported combination.

lib/Target/AMDGPU/AMDGPUISelLowering.cpp

SDValue AMDGPUTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
 
   case Intrinsic::AMDGPU_class:
     return DAG.getNode(AMDGPUISD::FP_CLASS, DL, VT,
                        Op.getOperand(1), Op.getOperand(2));
 
     case AMDGPUIntrinsic::AMDIL_exp: // Legacy name.
       return DAG.getNode(ISD::FEXP2, DL, VT, Op.getOperand(1));
 
-    case AMDGPUIntrinsic::AMDIL_round_nearest: // Legacy name.
-      return DAG.getNode(ISD::FRINT, DL, VT, Op.getOperand(1));
     case AMDGPUIntrinsic::AMDGPU_trunc: // Legacy name.
       return DAG.getNode(ISD::FTRUNC, DL, VT, Op.getOperand(1));
     case AMDGPUIntrinsic::AMDGPU_brev: // Legacy name
       return DAG.getNode(ISD::BITREVERSE, DL, VT, Op.getOperand(1));
   }
 }
 
 /// Linear Interpolation

lib/Target/AMDGPU/AMDGPUIntrinsics.td

let TargetPrefix = "AMDGPU", isTarget = 1 in {
   def int_AMDGPU_barrier_global  : Intrinsic<[], [], [IntrConvergent]>;
 }
 
 // Legacy names for compatibility.
 let TargetPrefix = "AMDIL", isTarget = 1 in {
   def int_AMDIL_fraction : Intrinsic<[llvm_anyfloat_ty], [LLVMMatchType<0>], [IntrNoMem]>;
   def int_AMDIL_clamp : Intrinsic<[llvm_anyfloat_ty], [LLVMMatchType<0>, LLVMMatchType<0>, LLVMMatchType<0>], [IntrNoMem]>;
   def int_AMDIL_exp : Intrinsic<[llvm_anyfloat_ty], [LLVMMatchType<0>], [IntrNoMem]>;
-  def int_AMDIL_round_nearest : Intrinsic<[llvm_anyfloat_ty], [LLVMMatchType<0>], [IntrNoMem]>;
 }
 
 let TargetPrefix = "TGSI", isTarget = 1 in {
 
   def int_TGSI_lit_z : Intrinsic<[llvm_float_ty], [llvm_float_ty, llvm_float_ty, llvm_float_ty],[IntrNoMem]>;
 }
 
 include "SIIntrinsics.td"

test/CodeGen/AMDGPU/llvm.rint.ll

 ; SI: v_rndne_f32_e32
 define void @rint_v4f32(<4 x float> addrspace(1)* %out, <4 x float> %in) {
 entry:
   %0 = call <4 x float> @llvm.rint.v4f32(<4 x float> %in) #0
   store <4 x float> %0, <4 x float> addrspace(1)* %out
   ret void
 }
 
-; FUNC-LABEL: {{^}}legacy_amdil_round_nearest_f32:
-; R600: RNDNE
-
-; SI: v_rndne_f32_e32
-define void @legacy_amdil_round_nearest_f32(float addrspace(1)* %out, float %in) {
-entry:
-  %0 = call float @llvm.AMDIL.round.nearest.f32(float %in) #0
-  store float %0, float addrspace(1)* %out
-  ret void
-}
-
-declare float @llvm.AMDIL.round.nearest.f32(float) #0
 declare float @llvm.rint.f32(float) #0
 declare <2 x float> @llvm.rint.v2f32(<2 x float>) #0
 declare <4 x float> @llvm.rint.v4f32(<4 x float>) #0
 
 attributes #0 = { nounwind readnone }