[llvm] [DirectX] Add support to lower LLVM intrinsics ceil, cos, fabs, and floor to DXIL Ops. (PR #80350)

[S. Bharadwaj Yadavalli via llvm-commits]

================
@@ -86,57 +106,85 @@ class dxil_op<string name, int code_id, dxil_class code_class, dxil_category op_
   let stats_group = _stats_group;
 }
 
-// The intrinsic which map directly to this dxil op.
+class dxil_op_min<string _name, int code_id, string _doc> : dxil_inst {
+  let name = _name;
+  let dxil_op = _name;
+  let dxil_opid = code_id;
+  let doc = _doc;
+}
+
+
+// LLVM intrinsic to map dxil op to.
 class dxil_map_intrinsic<Intrinsic llvm_intrinsic_> { Intrinsic llvm_intrinsic = llvm_intrinsic_; }
 
-def Sin : dxil_op<"Sin", 13, Unary, unary_float, "returns sine(theta) for theta in radians.",
-  "half;float;", "rn",
-  [
-    dxil_param<0, "$o", "", "operation result">,
-    dxil_param<1, "i32", "opcode", "DXIL opcode">,
-    dxil_param<2, "$o", "value", "input value">
-  ],
-  ["floats"]>,
-  dxil_map_intrinsic<int_sin>;
-
-def UMax :dxil_op< "UMax", 39,  Binary,  binary_uint, "unsigned integer maximum. UMax(a,b) = a > b ? a : b",
-    "i16;i32;i64;",  "rn",
-  [
+// Unary class DXIL Ops
+let op_class = unary, ops = [
+      dxil_param<0, "$o", "", "operation result">,
+      dxil_param<1, "i32", "opcode", "DXIL opcode">,
+      dxil_param<2, "$o", "value", "input value">
+    ], fn_attr = "rn" in {
+  // Unary half/float DXIL Ops
+  let category = unary_float, oload_types = "hf", stats_group = ["floats"] in {
----------------
bharadwajy wrote:

> `hf` means we need overloads for `half` and `float` right? Can we make that a list too? It would be nice to have the tablegen be less arcane and more explicit. I also question if these should be a string or a list of Type objects as defined above. The advantage of the latter is that we allow TableGen to verify the correctness.

Agreed. Will make the suggested changes. Thanks!


https://github.com/llvm/llvm-project/pull/80350