[LegalizeTypes][RISCV] Support scalable vectors in SplitVecRes_BITCAST/SplitVecOp_BITCAST. (llvm#88590)

Instead of bitcasting to a scalar integer, try to split the other vector
type that is legal. We already handle the case where both types need to
be split.

Fixes llvm#88576

Author: topperc

llvm/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp

@@ -1388,6 +1388,13 @@ void DAGTypeLegalizer::SplitVecRes_BITCAST(SDNode *N, SDValue &Lo,
     report_fatal_error("Scalarization of scalable vectors is not supported.");
   }
 
+  if (LoVT.isScalableVector()) {
+    auto [InLo, InHi] = DAG.SplitVectorOperand(N, 0);
+    Lo = DAG.getNode(ISD::BITCAST, dl, LoVT, InLo);
+    Hi = DAG.getNode(ISD::BITCAST, dl, HiVT, InHi);
+    return;
+  }
+
   // In the general case, convert the input to an integer and split it by hand.
   EVT LoIntVT = EVT::getIntegerVT(*DAG.getContext(), LoVT.getSizeInBits());
   EVT HiIntVT = EVT::getIntegerVT(*DAG.getContext(), HiVT.getSizeInBits());
@@ -3259,16 +3266,25 @@ SDValue DAGTypeLegalizer::SplitVecOp_BITCAST(SDNode *N) {
   // For example, i64 = BITCAST v4i16 on alpha.  Typically the vector will
   // end up being split all the way down to individual components.  Convert the
   // split pieces into integers and reassemble.
+  EVT ResVT = N->getValueType(0);
   SDValue Lo, Hi;
   GetSplitVector(N->getOperand(0), Lo, Hi);
+  SDLoc dl(N);
+
+  if (ResVT.isScalableVector()) {
+    auto [LoVT, HiVT] = DAG.GetSplitDestVTs(ResVT);
+    Lo = DAG.getNode(ISD::BITCAST, dl, LoVT, Lo);
+    Hi = DAG.getNode(ISD::BITCAST, dl, HiVT, Hi);
+    return DAG.getNode(ISD::CONCAT_VECTORS, dl, ResVT, Lo, Hi);
+  }
+
   Lo = BitConvertToInteger(Lo);
   Hi = BitConvertToInteger(Hi);
 
   if (DAG.getDataLayout().isBigEndian())
     std::swap(Lo, Hi);
 
-  return DAG.getNode(ISD::BITCAST, SDLoc(N), N->getValueType(0),
-                     JoinIntegers(Lo, Hi));
+  return DAG.getNode(ISD::BITCAST, dl, ResVT, JoinIntegers(Lo, Hi));
 }
 
 SDValue DAGTypeLegalizer::SplitVecOp_INSERT_SUBVECTOR(SDNode *N,

llvm/test/CodeGen/RISCV/rvv/pr88576.ll

@@ -0,0 +1,60 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 4
+; RUN: llc < %s -mtriple=riscv64 -mattr=+v | FileCheck %s
+
+define i1 @foo(<vscale x 16 x i8> %x, i64 %y) {
+; CHECK-LABEL: foo:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    csrr a1, vlenb
+; CHECK-NEXT:    slli a2, a1, 4
+; CHECK-NEXT:    addi a2, a2, -1
+; CHECK-NEXT:    bltu a0, a2, .LBB0_2
+; CHECK-NEXT:  # %bb.1:
+; CHECK-NEXT:    mv a0, a2
+; CHECK-NEXT:  .LBB0_2:
+; CHECK-NEXT:    addi sp, sp, -80
+; CHECK-NEXT:    .cfi_def_cfa_offset 80
+; CHECK-NEXT:    sd ra, 72(sp) # 8-byte Folded Spill
+; CHECK-NEXT:    sd s0, 64(sp) # 8-byte Folded Spill
+; CHECK-NEXT:    .cfi_offset ra, -8
+; CHECK-NEXT:    .cfi_offset s0, -16
+; CHECK-NEXT:    addi s0, sp, 80
+; CHECK-NEXT:    .cfi_def_cfa s0, 0
+; CHECK-NEXT:    csrr a2, vlenb
+; CHECK-NEXT:    slli a2, a2, 4
+; CHECK-NEXT:    sub sp, sp, a2
+; CHECK-NEXT:    andi sp, sp, -64
+; CHECK-NEXT:    addi a2, sp, 64
+; CHECK-NEXT:    add a0, a2, a0
+; CHECK-NEXT:    slli a1, a1, 3
+; CHECK-NEXT:    add a1, a2, a1
+; CHECK-NEXT:    vsetvli a3, zero, e8, m8, ta, ma
+; CHECK-NEXT:    vmv.v.i v16, 0
+; CHECK-NEXT:    vmv1r.v v0, v9
+; CHECK-NEXT:    vmerge.vim v24, v16, 1, v0
+; CHECK-NEXT:    vs8r.v v24, (a1)
+; CHECK-NEXT:    vmv1r.v v0, v8
+; CHECK-NEXT:    vmerge.vim v8, v16, 1, v0
+; CHECK-NEXT:    vs8r.v v8, (a2)
+; CHECK-NEXT:    lbu a0, 0(a0)
+; CHECK-NEXT:    addi sp, s0, -80
+; CHECK-NEXT:    ld ra, 72(sp) # 8-byte Folded Reload
+; CHECK-NEXT:    ld s0, 64(sp) # 8-byte Folded Reload
+; CHECK-NEXT:    addi sp, sp, 80
+; CHECK-NEXT:    ret
+  %a = bitcast <vscale x 16 x i8> %x to <vscale x 128 x i1>
+  %b = extractelement <vscale x 128 x i1> %a, i64 %y
+  ret i1 %b
+}
+
+define i8 @bar(<vscale x 128 x i1> %x, i64 %y) {
+; CHECK-LABEL: bar:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vmv1r.v v1, v8
+; CHECK-NEXT:    vsetivli zero, 1, e8, m2, ta, ma
+; CHECK-NEXT:    vslidedown.vx v8, v0, a0
+; CHECK-NEXT:    vmv.x.s a0, v8
+; CHECK-NEXT:    ret
+  %a = bitcast <vscale x 128 x i1> %x to <vscale x 16 x i8>
+  %b = extractelement <vscale x 16 x i8> %a, i64 %y
+  ret i8 %b
+}