[llvm] [RISCV] Add tests for vwaddu_wv+vabd(u) combine (PR #184962)

[Luke Lau via llvm-commits]

================
@@ -482,6 +482,112 @@ define <vscale x 4 x i32> @uabd_non_matching_promotion(<vscale x 4 x i8> %a, <vs
   ret <vscale x 4 x i32> %abs
 }
 
+define <vscale x 8 x i16> @vwabda(<vscale x 8 x i16> %acc ,<vscale x 8 x i8> %a, <vscale x 8 x i8> %b) {
+; CHECK-LABEL: vwabda:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vsetvli a0, zero, e8, m1, ta, ma
+; CHECK-NEXT:    vmin.vv v12, v10, v11
+; CHECK-NEXT:    vmax.vv v10, v10, v11
+; CHECK-NEXT:    vsub.vv v10, v10, v12
+; CHECK-NEXT:    vwaddu.wv v8, v8, v10
+; CHECK-NEXT:    ret
+;
+; ZVABD-LABEL: vwabda:
+; ZVABD:       # %bb.0:
+; ZVABD-NEXT:    vsetvli a0, zero, e8, m1, ta, ma
+; ZVABD-NEXT:    vabd.vv v10, v10, v11
+; ZVABD-NEXT:    vwaddu.wv v8, v8, v10
+; ZVABD-NEXT:    ret
+  %a.sext = sext <vscale x 8 x i8> %a to <vscale x 8 x i16>
+  %b.sext = sext <vscale x 8 x i8> %b to <vscale x 8 x i16>
+  %sub = sub <vscale x 8 x i16> %a.sext, %b.sext
+  %abs = call <vscale x 8 x i16> @llvm.abs.nxv8i16(<vscale x 8 x i16> %sub, i1 true)
+  %ret = add <vscale x 8 x i16> %acc, %abs
+  ret <vscale x 8 x i16> %ret
+}
+
+define <vscale x 8 x i16> @vwabdau(<vscale x 8 x i16> %acc ,<vscale x 8 x i8> %a, <vscale x 8 x i8> %b) {
+; CHECK-LABEL: vwabdau:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vsetvli a0, zero, e8, m1, ta, ma
+; CHECK-NEXT:    vminu.vv v12, v10, v11
+; CHECK-NEXT:    vmaxu.vv v10, v10, v11
+; CHECK-NEXT:    vsub.vv v10, v10, v12
+; CHECK-NEXT:    vwaddu.wv v8, v8, v10
+; CHECK-NEXT:    ret
+;
+; ZVABD-LABEL: vwabdau:
+; ZVABD:       # %bb.0:
+; ZVABD-NEXT:    vsetvli a0, zero, e8, m1, ta, ma
+; ZVABD-NEXT:    vabdu.vv v10, v10, v11
+; ZVABD-NEXT:    vwaddu.wv v8, v8, v10
+; ZVABD-NEXT:    ret
+  %a.zext = zext <vscale x 8 x i8> %a to <vscale x 8 x i16>
+  %b.zext = zext <vscale x 8 x i8> %b to <vscale x 8 x i16>
+  %sub = sub <vscale x 8 x i16> %a.zext, %b.zext
+  %abs = call <vscale x 8 x i16> @llvm.abs.nxv8i16(<vscale x 8 x i16> %sub, i1 true)
+  %ret = add <vscale x 8 x i16> %acc, %abs
+  ret <vscale x 8 x i16> %ret
+}
+
+define <vscale x 8 x i32> @vwabda_zext_diff(<vscale x 8 x i32> %acc ,<vscale x 8 x i8> %a, <vscale x 8 x i8> %b) {
+; CHECK-LABEL: vwabda_zext_diff:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vsetvli a0, zero, e8, m1, ta, ma
+; CHECK-NEXT:    vminu.vv v14, v12, v13
+; CHECK-NEXT:    vmaxu.vv v12, v12, v13
+; CHECK-NEXT:    vsub.vv v14, v12, v14
+; CHECK-NEXT:    vsetvli zero, zero, e16, m2, ta, ma
+; CHECK-NEXT:    vzext.vf2 v12, v14
+; CHECK-NEXT:    vwaddu.wv v8, v8, v12
+; CHECK-NEXT:    ret
+;
+; ZVABD-LABEL: vwabda_zext_diff:
+; ZVABD:       # %bb.0:
+; ZVABD-NEXT:    vsetvli a0, zero, e8, m1, ta, ma
+; ZVABD-NEXT:    vabdu.vv v14, v12, v13
+; ZVABD-NEXT:    vsetvli zero, zero, e16, m2, ta, ma
+; ZVABD-NEXT:    vzext.vf2 v12, v14
+; ZVABD-NEXT:    vwaddu.wv v8, v8, v12
+; ZVABD-NEXT:    ret
+  %a.zext = zext <vscale x 8 x i8> %a to <vscale x 8 x i16>
+  %b.zext = zext <vscale x 8 x i8> %b to <vscale x 8 x i16>
+  %sub = sub <vscale x 8 x i16> %a.zext, %b.zext
+  %abs = call <vscale x 8 x i16> @llvm.abs.nxv8i16(<vscale x 8 x i16> %sub, i1 true)
+  %abs.zext = zext <vscale x 8 x i16> %abs to <vscale x 8 x i32>
+  %ret = add <vscale x 8 x i32> %acc, %abs.zext
+  ret <vscale x 8 x i32> %ret
+}
+
+define <vscale x 8 x i32> @vwabdau_zext_diff(<vscale x 8 x i32> %acc ,<vscale x 8 x i8> %a, <vscale x 8 x i8> %b) {
----------------
lukel97 wrote:

I think this might have meant that the result of the llvm.abs is zexted, not the input arguments? 

https://github.com/llvm/llvm-project/pull/184962