[llvm] s390x: pattern match saturated truncation (PR #155377)

[Ulrich Weigand via llvm-commits]

================
@@ -424,6 +431,12 @@ let Predicates = [FeatureVector] in {
   defm VPKLSG : BinaryVRRbSPair<"vpklsg", 0xE795, int_s390_vpklsg, z_packls_cc,
                                 v128f, v128g, 3>;
 
+  // Pack
+  def VPK  : BinaryVRRcGeneric<"vpk", 0xE794>;
+  def VPKH : BinaryVRRc<"vpkh", 0xE794, z_pack, v128b, v128h, 1>;
+  def VPKF : BinaryVRRc<"vpkf", 0xE794, z_pack, v128h, v128f, 2>;
+  def VPKG : BinaryVRRc<"vpkg", 0xE794, z_pack, v128f, v128g, 3>;
----------------
uweigand wrote:

I don't think it should be necessary to move this - the tablegen logic should automatically prefer patterns that match more complex trees over simpler ones, independent of the order in the .td file.

https://github.com/llvm/llvm-project/pull/155377