[RISCV] Convert (splat_vector (load)) to vlse with 0 stride.

We already do this for splat nodes that carry a VL, but not for
splats that use VLMAX.

Reviewed By: frasercrmck

Differential Revision: https://reviews.llvm.org/D115483

diff --git a/llvm/lib/Target/RISCV/RISCVISelDAGToDAG.cpp b/llvm/lib/Target/RISCV/RISCVISelDAGToDAG.cpp
index 516e765..b24eb5f 100644 (file)
--- a/llvm/lib/Target/RISCV/RISCVISelDAGToDAG.cpp
+++ b/llvm/lib/Target/RISCV/RISCVISelDAGToDAG.cpp
@@ -1515,6 +1515,7 @@ void RISCVDAGToDAGISel::Select(SDNode *Node) {
     ReplaceNode(Node, Extract.getNode());
     return;
   }
+  case ISD::SPLAT_VECTOR:
   case RISCVISD::VMV_V_X_VL:
   case RISCVISD::VFMV_V_F_VL: {
     // Try to match splat of a scalar load to a strided load with stride of x0.
@@ -1531,7 +1532,10 @@ void RISCVDAGToDAGISel::Select(SDNode *Node) {
       break;
 
     SDValue VL;
-    selectVLOp(Node->getOperand(1), VL);
+    if (Node->getOpcode() == ISD::SPLAT_VECTOR)
+      VL = CurDAG->getTargetConstant(RISCV::VLMaxSentinel, DL, XLenVT);
+    else
+      selectVLOp(Node->getOperand(1), VL);
 
     unsigned Log2SEW = Log2_32(VT.getScalarSizeInBits());
     SDValue SEW = CurDAG->getTargetConstant(Log2SEW, DL, XLenVT);

diff --git a/llvm/test/CodeGen/RISCV/rvv/vsplats-fp.ll b/llvm/test/CodeGen/RISCV/rvv/vsplats-fp.ll
index d9b5583..a58822b 100644 (file)
--- a/llvm/test/CodeGen/RISCV/rvv/vsplats-fp.ll
+++ b/llvm/test/CodeGen/RISCV/rvv/vsplats-fp.ll
@@ -105,3 +105,22 @@ define <vscale x 8 x double> @vsplat_zero_nxv8f64() {
   %splat = shufflevector <vscale x 8 x double> %head, <vscale x 8 x double> undef, <vscale x 8 x i32> zeroinitializer
   ret <vscale x 8 x double> %splat
 }
+
+; Test that we fold this to a vlse with 0 stride.
+define <vscale x 8 x float> @vsplat_load_nxv8f32(float* %ptr) {
+; RV32V-LABEL: vsplat_load_nxv8f32:
+; RV32V:       # %bb.0:
+; RV32V-NEXT:    vsetvli a1, zero, e32, m4, ta, mu
+; RV32V-NEXT:    vlse32.v v8, (a0), zero
+; RV32V-NEXT:    ret
+;
+; RV64V-LABEL: vsplat_load_nxv8f32:
+; RV64V:       # %bb.0:
+; RV64V-NEXT:    vsetvli a1, zero, e32, m4, ta, mu
+; RV64V-NEXT:    vlse32.v v8, (a0), zero
+; RV64V-NEXT:    ret
+  %f = load float, float* %ptr
+  %head = insertelement <vscale x 8 x float> undef, float %f, i32 0
+  %splat = shufflevector <vscale x 8 x float> %head, <vscale x 8 x float> undef, <vscale x 8 x i32> zeroinitializer
+  ret <vscale x 8 x float> %splat
+}