[llvm-branch-commits] [llvm] [AArch64] Split large loop dependence masks (PR #153187)

[Sander de Smalen via llvm-branch-commits]

================
@@ -5248,49 +5248,94 @@ AArch64TargetLowering::LowerLOOP_DEPENDENCE_MASK(SDValue Op,
                                                  SelectionDAG &DAG) const {
   SDLoc DL(Op);
   uint64_t EltSize = Op.getConstantOperandVal(2);
-  EVT VT = Op.getValueType();
+  EVT FullVT = Op.getValueType();
+  unsigned NumElements = FullVT.getVectorMinNumElements();
+  unsigned NumSplits = 0;
+  EVT EltVT;
   switch (EltSize) {
   case 1:
-    if (VT != MVT::v16i8 && VT != MVT::nxv16i1)
-      return SDValue();
+    EltVT = MVT::i8;
     break;
   case 2:
-    if (VT != MVT::v8i8 && VT != MVT::nxv8i1)
-      return SDValue();
+    if (NumElements >= 16)
----------------
sdesmalen-arm wrote:

When the number of elements is smaller, not returning `SDValue()` results in selectiondag failures, e.g.

```
define <vscale x 4 x i1> @whilewr_64_split(ptr %a, ptr %b) {
entry:
  %0 = call <vscale x 4 x i1> @llvm.loop.dependence.war.mask.nxv4i1(ptr %a, ptr %b, i64 1)
  ret <vscale x 4 x i1> %0
}
```

There seems to be missing test-coverage for that.

https://github.com/llvm/llvm-project/pull/153187