AMDGPU: Custom lower v4i16/v4f16 vector operations

Avoids stack access.

Also handle extract hi elt pattern from truncate + shift
to avoid a couple test regressions.

llvm-svn: 332453

1 files changed, 22 insertions, 0 deletions

diff --git a/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp b/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp
index d00727bf314..988554621c5 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp
@@ -3144,6 +3144,28 @@ SDValue AMDGPUTargetLowering::performTruncateCombine(
     }
   }
 
+  // Equivalent of above for accessing the high element of a vector as an
+  // integer operation.
+  // trunc (srl (bitcast (build_vector x, y))), 16 -> trunc (bitcast y)
+  if (Src.getOpcode() == ISD::SRL) {
+    if (auto K = isConstOrConstSplat(Src.getOperand(1))) {
+      if (2 * K->getZExtValue() == Src.getValueType().getScalarSizeInBits()) {
+        SDValue BV = stripBitcast(Src.getOperand(0));
+        if (BV.getOpcode() == ISD::BUILD_VECTOR &&
+            BV.getValueType().getVectorNumElements() == 2) {
+          SDValue SrcElt = BV.getOperand(1);
+          EVT SrcEltVT = SrcElt.getValueType();
+          if (SrcEltVT.isFloatingPoint()) {
+            SrcElt = DAG.getNode(ISD::BITCAST, SL,
+                                 SrcEltVT.changeTypeToInteger(), SrcElt);
+          }
+
+          return DAG.getNode(ISD::TRUNCATE, SL, VT, SrcElt);
+        }
+      }
+    }
+  }
+
   // Partially shrink 64-bit shifts to 32-bit if reduced to 16-bit.
   //
   // i16 (trunc (srl i64:x, K)), K <= 16 ->