[SelectionDAG] Add BITCAST handling to ComputeNumSignBits for splatted sign bits.

For cases where we are BITCASTing to vectors of smaller elements, then if the entire source was a splatted sign (src's NumSignBits == SrcBitWidth) we can say that the dst's NumSignBit == DstBitWidth, as we're just splitting those sign bits across multiple elements.

We could generalize this but at the moment the only use case I have is to peek through bitcasts to vector comparison results.

Differential Revision: https://reviews.llvm.org/D37849

llvm-svn: 313543

1 files changed, 24 insertions, 0 deletions

diff --git a/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp b/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
index 85ef6878490..4942f745f6c 100644
--- a/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
@@ -3035,6 +3035,30 @@ unsigned SelectionDAG::ComputeNumSignBits(SDValue Op, const APInt &DemandedElts,
     return Tmp;
   }
 
+  case ISD::BITCAST: {
+    SDValue N0 = Op.getOperand(0);
+    unsigned SrcBits = N0.getScalarValueSizeInBits();
+
+    // Ignore bitcasts from floating point.
+    if (!N0.getValueType().isInteger())
+      break;
+
+    // Fast handling of 'identity' bitcasts.
+    if (VTBits == SrcBits)
+      return ComputeNumSignBits(N0, DemandedElts, Depth + 1);
+
+    // Bitcast 'large element' scalar/vector to 'small element' vector.
+    // TODO: Handle cases other than 'sign splat' when we have a use case.
+    // Requires handling of DemandedElts and Endianness.
+    if ((SrcBits % VTBits) == 0) {
+      assert(Op.getValueType().isVector() && "Expected bitcast to vector");
+      Tmp = ComputeNumSignBits(N0, Depth + 1);
+      if (Tmp == SrcBits)
+        return VTBits;
+    }
+    break;
+  }
+
   case ISD::SIGN_EXTEND:
   case ISD::SIGN_EXTEND_VECTOR_INREG:
     Tmp = VTBits - Op.getOperand(0).getScalarValueSizeInBits();