[llvm] [SystemZ] Improve handling of memmoves. (PR #196285)

[Ulrich Weigand via llvm-commits]

================
@@ -1470,6 +1470,14 @@ bool SystemZTargetLowering::findOptimalMemOpLowering(
   assert(Limit != ~0U &&
          "Expected EmitTargetCodeForMemXXX() to handle AlwaysInline cases.");
 
+  if (Op.isMemmove()) {
+    if (Op.size() >= 16 &&
+        (!Op.isAligned(Align(8)) || (Op.size() >= 25 && Op.size() <= 31)))
+      return false;
+    return TargetLowering::findOptimalMemOpLowering(
+        Context, MemOps, Limit, Op, DstAS, SrcAS, FuncAttributes, LargestVT);
+  }
----------------
uweigand wrote:

I don't really like the duplication of logic here.   Could we rewrite this by removing this block, and changing the subsequent blocks to something like:
```
  const int MVCFastLen = 16;
  // Use MVC up to 16 bytes for memcpy.  Small memset uses STC/MVI for first byte.
  if (Op.isMemcpy() && Op.size() <= MVCFastLen)
    return false;
  if (Op.isMemset() && Op.size() - 1 <= MVCFastLen)
    return false;

  // Avoid unaligned VL/VST:s.
  if ((Op.size() >= 16 && !Op.isAligned(Align(8))) ||
      (Op.size() >= 25 && Op.size() <= 31))
    return false;
```


https://github.com/llvm/llvm-project/pull/196285