[AMDGPU] should expand ROTL i16 to shifts.

[alex-t]

Summary

Instruction combining pass turns library rotl implementation to llvm.fshl.i16.
In the selection dag the intrinsic is turned to ISD::ROTL node that cannot be selected.
Need to expand it to shifts again.

[rampitec]

Needs test.

llvm/lib/Target/AMDGPU/SIISelLowering.cpp

548–549

What about rotr?

[arsenm]

Needs test

[alex-t]

Tests added. ROTR case added.

[rampitec]

LGTM

[arsenm]

Should go into existing fshr and fshr tests

[alex-t]

tests moved to existing rotl/rotr tests

llvm/lib/Target/AMDGPU/SIISelLowering.cpp

if (Subtarget->has16BitInsts()) {
     setOperationAction(ISD::SMIN, MVT::i16, Legal);
     setOperationAction(ISD::SMAX, MVT::i16, Legal);
 
     setOperationAction(ISD::UMIN, MVT::i16, Legal);
     setOperationAction(ISD::UMAX, MVT::i16, Legal);
 
     setOperationAction(ISD::SIGN_EXTEND, MVT::i16, Promote);
     AddPromotedToType(ISD::SIGN_EXTEND, MVT::i16, MVT::i32);
 
-    setOperationAction(ISD::ROTR, MVT::i16, Promote);
-    setOperationAction(ISD::ROTL, MVT::i16, Promote);
+    setOperationAction(ISD::ROTR, MVT::i16, Expand);
+    setOperationAction(ISD::ROTL, MVT::i16, Expand);
 
     setOperationAction(ISD::SDIV, MVT::i16, Promote);
     setOperationAction(ISD::UDIV, MVT::i16, Promote);
     setOperationAction(ISD::SREM, MVT::i16, Promote);
     setOperationAction(ISD::UREM, MVT::i16, Promote);
     setOperationAction(ISD::UADDSAT, MVT::i16, Legal);
     setOperationAction(ISD::USUBSAT, MVT::i16, Legal);
 

llvm/test/CodeGen/AMDGPU/rotl.ll

 entry:
   %0 = shl <4 x i32> %x, %y
   %1 = sub <4 x i32> <i32 32, i32 32, i32 32, i32 32>, %y
   %2 = lshr <4 x i32> %x, %1
   %3 = or <4 x i32> %0, %2
   store <4 x i32> %3, <4 x i32> addrspace(1)* %in
   ret void
 }
+
+; GCN-LABEL: @test_rotl_i16
+; GCN: global_load_ushort [[X:v[0-9]+]]
+; GCN: global_load_ushort [[D:v[0-9]+]]
+; GCN: v_sub_nc_u16_e64 [[NX:v[0-9]+]], 0, [[X]]
+; GCN: v_and_b32_e32 [[XAND:v[0-9]+]], 15, [[X]]
+; GCN: v_and_b32_e32 [[NXAND:v[0-9]+]], 15, [[NX]]
+; GCN: v_lshlrev_b16_e64 [[LO:v[0-9]+]], [[XAND]], [[D]]
+; GCN: v_lshrrev_b16_e64 [[HI:v[0-9]+]], [[NXAND]], [[D]]
+; GCN: v_or_b32_e32 [[RES:v[0-9]+]], [[LO]], [[HI]]
+; GCN: global_store_short v{{\[[0-9]+:[0-9]+\]}}, [[RES]]
+
+declare i16 @llvm.fshl.i16(i16, i16, i16)
+
+define void @test_rotl_i16(i16 addrspace(1)* nocapture readonly %sourceA, i16 addrspace(1)* nocapture readonly %sourceB, i16 addrspace(1)* nocapture %destValues) {
+entry:
+  %arrayidx = getelementptr inbounds i16, i16 addrspace(1)* %sourceA, i64 16
+  %a = load i16, i16 addrspace(1)* %arrayidx
+  %arrayidx2 = getelementptr inbounds i16, i16 addrspace(1)* %sourceB, i64 24
+  %b = load i16, i16 addrspace(1)* %arrayidx2
+  %c = tail call i16 @llvm.fshl.i16(i16 %a, i16 %a, i16 %b)
+  %arrayidx5 = getelementptr inbounds i16, i16 addrspace(1)* %destValues, i64 4
+  store i16 %c, i16 addrspace(1)* %arrayidx5
+  ret void
+}

llvm/test/CodeGen/AMDGPU/rotr.ll

 entry:
   %tmp0 = sub <4 x i32> <i32 32, i32 32, i32 32, i32 32>, %y
   %tmp1 = shl <4 x i32> %x, %tmp0
   %tmp2 = lshr <4 x i32> %x, %y
   %tmp3 = or <4 x i32> %tmp1, %tmp2
   store <4 x i32> %tmp3, <4 x i32> addrspace(1)* %in
   ret void
 }
+
+; GCN-LABEL: @test_rotr_i16
+; GCN: global_load_ushort [[X:v[0-9]+]]
+; GCN: global_load_ushort [[D:v[0-9]+]]
+; GCN: v_sub_nc_u16_e64 [[NX:v[0-9]+]], 0, [[X]]
+; GCN: v_and_b32_e32 [[XAND:v[0-9]+]], 15, [[X]]
+; GCN: v_and_b32_e32 [[NXAND:v[0-9]+]], 15, [[NX]]
+; GCN: v_lshrrev_b16_e64 [[LO:v[0-9]+]], [[XAND]], [[D]]
+; GCN: v_lshlrev_b16_e64 [[HI:v[0-9]+]], [[NXAND]], [[D]]
+; GCN: v_or_b32_e32 [[RES:v[0-9]+]], [[LO]], [[HI]]
+; GCN: global_store_short v{{\[[0-9]+:[0-9]+\]}}, [[RES]]
+
+declare i16 @llvm.fshr.i16(i16, i16, i16)
+
+define void @test_rotr_i16(i16 addrspace(1)* nocapture readonly %sourceA, i16 addrspace(1)* nocapture readonly %sourceB, i16 addrspace(1)* nocapture %destValues) {
+entry:
+  %arrayidx = getelementptr inbounds i16, i16 addrspace(1)* %sourceA, i64 16
+  %a = load i16, i16 addrspace(1)* %arrayidx
+  %arrayidx2 = getelementptr inbounds i16, i16 addrspace(1)* %sourceB, i64 24
+  %b = load i16, i16 addrspace(1)* %arrayidx2
+  %c = tail call i16 @llvm.fshr.i16(i16 %a, i16 %a, i16 %b)
+  %arrayidx5 = getelementptr inbounds i16, i16 addrspace(1)* %destValues, i64 4
+  store i16 %c, i16 addrspace(1)* %arrayidx5
+  ret void
+}