Fixed the DAG combiner to better handle the folding of AND nodes for vector types...

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@163203 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
index ee6c2a39f3ed71c2b351ce6abc4e272afdc009f8..45c38968ea22d9010e55d0c7793e8ceec8d9aa6f 100644 (file)
--- a/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ b/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -2496,8 +2496,18 @@ SDValue DAGCombiner::visitAND(SDNode *N) {
         // lanes of the constant together.
         EVT VT = Vector->getValueType(0);
         unsigned BitWidth = VT.getVectorElementType().getSizeInBits();
+
+        // If the splat value has been compressed to a bitlength lower
+        // than the size of the vector lane, we need to re-expand it to
+        // the lane size.
+        if (BitWidth > SplatBitSize)
+          for (SplatValue = SplatValue.zextOrTrunc(BitWidth);
+               SplatBitSize < BitWidth;
+               SplatBitSize = SplatBitSize * 2)
+            SplatValue |= SplatValue.shl(SplatBitSize);
+
         Constant = APInt::getAllOnesValue(BitWidth);
-        for (unsigned i = 0, n = VT.getVectorNumElements(); i < n; ++i)
+        for (unsigned i = 0, n = SplatBitSize/BitWidth; i < n; ++i)
           Constant &= SplatValue.lshr(i*BitWidth).zextOrTrunc(BitWidth);
       }
     }

diff --git a/test/CodeGen/ARM/vector-extend-narrow.ll b/test/CodeGen/ARM/vector-extend-narrow.ll
index 8fd3db29197e878d337cd22d94a2b1ebb4901bf3..22af7976212802accc8386fbf99a9273e44206ff 100644 (file)
--- a/test/CodeGen/ARM/vector-extend-narrow.ll
+++ b/test/CodeGen/ARM/vector-extend-narrow.ll
@@ -62,3 +62,14 @@ define <4 x i8> @i(<4 x i8>* %x) {
   %2 = sdiv <4 x i8> zeroinitializer, %1
   ret <4 x i8> %2
 }
+; CHECK: j:
+define <4 x i32> @j(<4 x i8>* %in) nounwind {
+  ; CHECK: vld1
+  ; CHECK: vmovl.u8
+  ; CHECK: vmovl.u16
+  ; CHECK-NOT: vand
+  %1 = load <4 x i8>* %in, align 4
+  %2 = zext <4 x i8> %1 to <4 x i32>
+  ret <4 x i32> %2
+}
+