[llvm] [SelectionDAG] Improve isGuaranteedNotToBeUndefOrPoison (PR #147019)

[Björn Pettersson via llvm-commits]

================
@@ -5456,6 +5456,85 @@ bool SelectionDAG::isGuaranteedNotToBeUndefOrPoison(SDValue Op,
     }
     return true;
 
+  case ISD::EXTRACT_SUBVECTOR: {
+    SDValue Src = Op.getOperand(0);
+    if (Src.getValueType().isScalableVector())
+      break;
+    uint64_t Idx = Op.getConstantOperandVal(1);
+    unsigned NumSrcElts = Src.getValueType().getVectorNumElements();
+    APInt DemandedSrcElts = DemandedElts.zext(NumSrcElts).shl(Idx);
+    return isGuaranteedNotToBeUndefOrPoison(Src, DemandedSrcElts, PoisonOnly,
+                                            Depth + 1);
+  }
+
+  case ISD::INSERT_SUBVECTOR: {
+    if (Op.getValueType().isScalableVector())
+      break;
+    SDValue Src = Op.getOperand(0);
+    SDValue Sub = Op.getOperand(1);
+    uint64_t Idx = Op.getConstantOperandVal(2);
+    unsigned NumSubElts = Sub.getValueType().getVectorNumElements();
+    APInt DemandedSubElts = DemandedElts.extractBits(NumSubElts, Idx);
+    APInt DemandedSrcElts = DemandedElts;
+    DemandedSrcElts.clearBits(Idx, Idx + NumSubElts);
+
+    if (!!DemandedSubElts && !isGuaranteedNotToBeUndefOrPoison(
+                                 Sub, DemandedSubElts, PoisonOnly, Depth + 1))
+      return false;
+    if (!!DemandedSrcElts && !isGuaranteedNotToBeUndefOrPoison(
+                                 Src, DemandedSrcElts, PoisonOnly, Depth + 1))
+      return false;
+    return true;
+  }
+
+  case ISD::EXTRACT_VECTOR_ELT: {
+    SDValue Src = Op.getOperand(0);
+    auto *IndexC = dyn_cast<ConstantSDNode>(Op.getOperand(1));
+    EVT SrcVT = Src.getValueType();
+    if (SrcVT.isFixedLengthVector() && IndexC &&
+        IndexC->getAPIntValue().ult(SrcVT.getVectorNumElements())) {
+      APInt DemandedSrcElts = APInt::getOneBitSet(SrcVT.getVectorNumElements(),
+                                                  IndexC->getZExtValue());
+      return isGuaranteedNotToBeUndefOrPoison(Src, DemandedSrcElts, PoisonOnly,
+                                              Depth + 1);
+    }
+    break;
+  }
+
+  case ISD::INSERT_VECTOR_ELT: {
+    SDValue InVec = Op.getOperand(0);
+    SDValue InVal = Op.getOperand(1);
+    SDValue EltNo = Op.getOperand(2);
+    EVT VT = InVec.getValueType();
+    auto *IndexC = dyn_cast<ConstantSDNode>(EltNo);
+    if (IndexC && VT.isFixedLengthVector() &&
+        IndexC->getZExtValue() < VT.getVectorNumElements()) {
+      if (DemandedElts[IndexC->getZExtValue()] &&
+          !isGuaranteedNotToBeUndefOrPoison(InVal, PoisonOnly, Depth + 1))
+        return false;
+      APInt InVecDemandedElts = DemandedElts;
+      InVecDemandedElts.clearBit(IndexC->getZExtValue());
+      if (!!InVecDemandedElts &&
+          !isGuaranteedNotToBeUndefOrPoison(
+              peekThroughInsertVectorElt(InVec, InVecDemandedElts),
+              InVecDemandedElts, PoisonOnly, Depth + 1))
+        return false;
+      return true;
+    }
+    break;
+  }
+
+  case ISD::SCALAR_TO_VECTOR:
+    // Check upper (known undef) elements.
+    if (DemandedElts.ugt(1) && !PoisonOnly)
+      return false;
+    // Check element zero.
+    if (DemandedElts[0] && !isGuaranteedNotToBeUndefOrPoison(Op.getOperand(0),
+                                                             PoisonOnly,
+                                                             Depth + 1))
+      return false;
+    return true;
----------------
bjope wrote:

Consider PoisonOnly==true and only the upper undef elements are demanded. We should be allowed to return true indicating that the upper elements aren't poison. Although not sure if that would be a very rare situation.

There is also the odd case when DemandedElts==0 for which we now return true. I'm basically same logic as for operations like INSERT_SUBVECTOR and VECTOR_SHUFFLE here. That we first check !isGuaranteedNot for values that map to the demanded elements. If we passed all those checks then isGuaranteedNot should be true for all demanded elements and we return true. And in case DemandedElts==0 we also end up returning true.

https://github.com/llvm/llvm-project/pull/147019