[llvm] [SPIRV] Use `Op[S|U]Dot` when possible for integer dot product (PR #115095)

[Finn Plummer via llvm-commits]

================
@@ -73,16 +127,21 @@ define noundef i64 @dot_uint64_t4(<2 x i64> noundef %a, <2 x i64> noundef %b) {
 entry:
 ; CHECK: %[[#arg0:]] = OpFunctionParameter %[[#vec2_int_64]]
 ; CHECK: %[[#arg1:]] = OpFunctionParameter %[[#vec2_int_64]]
-; CHECK: %[[#mul_vec:]] = OpIMul %[[#vec2_int_64]] %[[#arg0]] %[[#arg1]]
-; CHECK: %[[#elt0:]] = OpCompositeExtract %[[#int_64]] %[[#mul_vec]] 0
-; CHECK: %[[#elt1:]] = OpCompositeExtract %[[#int_64]] %[[#mul_vec]] 1
-; CHECK: %[[#sum0:]] = OpIAdd %[[#int_64]] %[[#elt0]] %[[#elt1]]
+
+; CHECK-DOT: %[[#dot:]] = OpUDot %[[#int_64]] %[[#arg0]] %[[#arg1]]
+
+; CHECK-EXP: %[[#mul_vec:]] = OpIMul %[[#vec2_int_64]] %[[#arg0]] %[[#arg1]]
+; CHECK-EXP: %[[#elt0:]] = OpCompositeExtract %[[#int_64]] %[[#mul_vec]] 0
+; CHECK-EXP: %[[#elt1:]] = OpCompositeExtract %[[#int_64]] %[[#mul_vec]] 1
+; CHECK-EXP: %[[#sum0:]] = OpIAdd %[[#int_64]] %[[#elt0]] %[[#elt1]]
   %dot = call i64 @llvm.spv.udot.v2i64(<2 x i64> %a, <2 x i64> %b)
   ret i64 %dot
 }
 
+declare i8 @llvm.spv.sdot.v4i8(<4 x i8>, <4 x i8>)
 declare i16 @llvm.spv.sdot.v2i16(<2 x i16>, <2 x i16>)
 declare i32 @llvm.spv.sdot.v4i32(<4 x i32>, <4 x i32>)
+declare i8 @llvm.spv.udot.v4i8(<4 x i8>, <4 x i8>)
----------------
inbelic wrote:

Do you feel strongly about this? I think in all the testcases I have seen there is an incomplete declaration list, so I would vote in being consistent with that.

https://github.com/llvm/llvm-project/pull/115095