AMDGPU : Additional pattern for median3 matching

[aakanksha555]

Summary

Found a missing pattern for v_med3_i16 matching
min(max(a, b), max(min(a, b), c))
Added tests to support this pattern

[arsenm]

LGTM

lib/Target/AMDGPU/SIInstructions.td

1667–1668

Spaces before //

lib/Target/AMDGPU/SIInstructions.td

class FP16Med3Pat<ValueType vt,
   (fmaxnum_like (fminnum_like_oneuse (VOP3Mods_nnan vt:$src0, i32:$src0_mods),
                                      (VOP3Mods_nnan vt:$src1, i32:$src1_mods)),
            (fminnum_like_oneuse (fmaxnum_like_oneuse (VOP3Mods_nnan vt:$src0, i32:$src0_mods),
                                                      (VOP3Mods_nnan vt:$src1, i32:$src1_mods)),
                            (vt (VOP3Mods_nnan vt:$src2, i32:$src2_mods)))),
   (med3Inst $src0_mods, $src0, $src1_mods, $src1, $src2_mods, $src2, DSTCLAMP.NONE)
 >;
 
-class Int16Med3Pat<Instruction med3Inst,
+multiclass Int16Med3Pat<Instruction med3Inst,
+                   SDPatternOperator min,
                    SDPatternOperator max,
                    SDPatternOperator max_oneuse,
                    SDPatternOperator min_oneuse,
-                   ValueType vt = i32> : GCNPat<
+                   ValueType vt = i16> {
+  // This matches 16 permutations of
+  // max(min(x, y), min(max(x, y), z))
+  def : GCNPat <
   (max (min_oneuse vt:$src0, vt:$src1),
        (min_oneuse (max_oneuse vt:$src0, vt:$src1), vt:$src2)),
   (med3Inst SRCMODS.NONE, $src0, SRCMODS.NONE, $src1, SRCMODS.NONE, $src2, DSTCLAMP.NONE)
 >;
 
+  //This matches 16 permutations of
+  //min(max(a, b), max(min(a, b), c))
+  def : GCNPat <
+  (min (max_oneuse vt:$src0, vt:$src1),
+      (max_oneuse (min_oneuse vt:$src0, vt:$src1), vt:$src2)),
+  (med3Inst SRCMODS.NONE, $src0, SRCMODS.NONE, $src1, SRCMODS.NONE, $src2, DSTCLAMP.NONE)
+>;
+}
+
 def : FPMed3Pat<f32, V_MED3_F32>;
 
 let OtherPredicates = [isGFX9] in {
 def : FP16Med3Pat<f16, V_MED3_F16>;
-def : Int16Med3Pat<V_MED3_I16, smax, smax_oneuse, smin_oneuse, i16>;
-def : Int16Med3Pat<V_MED3_U16, umax, umax_oneuse, umin_oneuse, i16>;
+defm : Int16Med3Pat<V_MED3_I16, smin, smax, smax_oneuse, smin_oneuse>;
+defm : Int16Med3Pat<V_MED3_U16, umin, umax, umax_oneuse, umin_oneuse>;
 } // End Predicates = [isGFX9]

test/CodeGen/AMDGPU/smed3.ll

bb:
   %tmp0 = call i16 @smin16(i16 %x, i16 %y)
   %tmp1 = call i16 @smax16(i16 %x, i16 %y)
   %tmp2 = call i16 @smin16(i16 %tmp1, i16 %z)
   %tmp3 = call i16 @smax16(i16 %tmp0, i16 %tmp2)
   store i16 %tmp3, i16 addrspace(1)* %out.gep
   ret void
 }
 
+; GCN-LABEL: {{^}}v_test_smed3_i16_pat_1:
+; GFX9: v_med3_i16 v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}}
+
+define amdgpu_kernel void @v_test_smed3_i16_pat_1(i16 addrspace(1)* %arg, i16 addrspace(1)* %out, i16 addrspace(1)* %a.ptr) #1 {
+bb:
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %gep0 = getelementptr inbounds i16, i16 addrspace(1)* %a.ptr, i32 %tid
+  %gep1 = getelementptr inbounds i16, i16 addrspace(1)* %gep0, i32 3
+  %gep2 = getelementptr inbounds i16, i16 addrspace(1)* %gep0, i32 8
+  %out.gep = getelementptr inbounds i16, i16 addrspace(1)* %out, i32 %tid
+  %x = load i16, i16 addrspace(1)* %gep0
+  %y = load i16, i16 addrspace(1)* %gep1
+  %z = load i16, i16 addrspace(1)* %gep2
+
+  %tmp0 = call i16 @smin16(i16 %x, i16 %y)
+  %tmp1 = call i16 @smax16(i16 %x, i16 %y)
+  %tmp2 = call i16 @smax16(i16 %tmp0, i16 %z)
+  %tmp3 = call i16 @smin16(i16 %tmp1, i16 %tmp2)
+  store i16 %tmp3, i16 addrspace(1)* %out.gep
+  ret void
+}
+
 attributes #0 = { nounwind readnone }
 attributes #1 = { nounwind }
 attributes #2 = { nounwind readnone alwaysinline }

test/CodeGen/AMDGPU/umed3.ll

bb:
   %tmp0 = call i16 @umin16(i16 %x, i16 %y)
   %tmp1 = call i16 @umax16(i16 %x, i16 %y)
   %tmp2 = call i16 @umin16(i16 %tmp1, i16 %z)
   %tmp3 = call i16 @umax16(i16 %tmp0, i16 %tmp2)
   store i16 %tmp3, i16 addrspace(1)* %out.gep
   ret void
 }
 
+; GCN-LABEL: {{^}}v_test_umed3_i16_pat_1:
+; GFX9: v_med3_u16 v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}}
+define amdgpu_kernel void @v_test_umed3_i16_pat_1(i16 addrspace(1)* %arg, i16 addrspace(1)* %out, i16 addrspace(1)* %a.ptr) #1 {
+bb:
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %gep0 = getelementptr inbounds i16, i16 addrspace(1)* %a.ptr, i32 %tid
+  %gep1 = getelementptr inbounds i16, i16 addrspace(1)* %gep0, i32 3
+  %gep2 = getelementptr inbounds i16, i16 addrspace(1)* %gep0, i32 8
+  %out.gep = getelementptr inbounds i16, i16 addrspace(1)* %out, i32 %tid
+  %x = load i16, i16 addrspace(1)* %gep0
+  %y = load i16, i16 addrspace(1)* %gep1
+  %z = load i16, i16 addrspace(1)* %gep2
+
+  %tmp0 = call i16 @umin16(i16 %x, i16 %y)
+  %tmp1 = call i16 @umax16(i16 %x, i16 %y)
+  %tmp2 = call i16 @umax16(i16 %tmp0, i16 %z)
+  %tmp3 = call i16 @umin16(i16 %tmp1, i16 %tmp2)
+  store i16 %tmp3, i16 addrspace(1)* %out.gep
+  ret void
+}
+
 attributes #0 = { nounwind readnone }
 attributes #1 = { nounwind }
 attributes #2 = { nounwind readnone alwaysinline }