[RISCV] Rewrite deinterleave load as vlse optimization as DAG combine (#150049)

This reworks an existing optimization on the fixed vector (shuffle
based) deinterleave lowering into a DAG combine. This has the effect of
making it kick in much more widely - in particular on the deinterleave
intrinsic (i.e. scalable) path, deinterleaveN (without load) lowering,
but also the intrinsic lowering paths.

Author: preames

llvm/lib/Target/RISCV/RISCVISelLowering.cpp

@@ -20843,6 +20843,62 @@ SDValue RISCVTargetLowering::PerformDAGCombine(SDNode *N,
     }
     break;
   }
+  case RISCVISD::TUPLE_EXTRACT: {
+    EVT VT = N->getValueType(0);
+    SDValue Tuple = N->getOperand(0);
+    unsigned Idx = N->getConstantOperandVal(1);
+    if (!Tuple.hasOneUse() || Tuple.getOpcode() != ISD::INTRINSIC_W_CHAIN)
+      break;
+
+    unsigned NF = 0;
+    switch (Tuple.getConstantOperandVal(1)) {
+    default:
+      break;
+    case Intrinsic::riscv_vlseg2_mask:
+    case Intrinsic::riscv_vlseg3_mask:
+    case Intrinsic::riscv_vlseg4_mask:
+    case Intrinsic::riscv_vlseg5_mask:
+    case Intrinsic::riscv_vlseg6_mask:
+    case Intrinsic::riscv_vlseg7_mask:
+    case Intrinsic::riscv_vlseg8_mask:
+      NF = Tuple.getValueType().getRISCVVectorTupleNumFields();
+      break;
+    }
+
+    if (!NF || Subtarget.hasOptimizedSegmentLoadStore(NF))
+      break;
+
+    unsigned SEW = VT.getScalarSizeInBits();
+    assert(Log2_64(SEW) == Tuple.getConstantOperandVal(7) &&
+           "Type mismatch without bitcast?");
+    unsigned Stride = SEW / 8 * NF;
+    unsigned Offset = SEW / 8 * Idx;
+
+    SDValue Ops[] = {
+        /*Chain=*/Tuple.getOperand(0),
+        /*IntID=*/DAG.getTargetConstant(Intrinsic::riscv_vlse_mask, DL, XLenVT),
+        /*Passthru=*/Tuple.getOperand(2),
+        /*Ptr=*/
+        DAG.getNode(ISD::ADD, DL, XLenVT, Tuple.getOperand(3),
+                    DAG.getConstant(Offset, DL, XLenVT)),
+        /*Stride=*/DAG.getConstant(Stride, DL, XLenVT),
+        /*Mask=*/Tuple.getOperand(4),
+        /*VL=*/Tuple.getOperand(5),
+        /*Policy=*/Tuple.getOperand(6)};
+
+    auto TupleMemSD = cast<MemIntrinsicSDNode>(Tuple);
+    // Match getTgtMemIntrinsic for non-unit stride case
+    EVT MemVT = TupleMemSD->getMemoryVT().getScalarType();
+    MachineFunction &MF = DAG.getMachineFunction();
+    MachineMemOperand *MMO = MF.getMachineMemOperand(
+        TupleMemSD->getMemOperand(), Offset, MemoryLocation::UnknownSize);
+
+    SDVTList VTs = DAG.getVTList({VT, MVT::Other});
+    SDValue Result = DAG.getMemIntrinsicNode(ISD::INTRINSIC_W_CHAIN, DL, VTs,
+                                             Ops, MemVT, MMO);
+    DAG.ReplaceAllUsesOfValueWith(Tuple.getValue(1), Result.getValue(1));
+    return Result.getValue(0);
+  }
   }
 
   return SDValue();

llvm/lib/Target/RISCV/RISCVInterleavedAccess.cpp

@@ -216,29 +216,6 @@ bool RISCVTargetLowering::lowerInterleavedLoad(
   if (!isLegalInterleavedAccessType(VTy, Factor, Alignment, AS, DL))
     return false;
 
-  // If the segment load is going to be performed segment at a time anyways
-  // and there's only one element used, use a strided load instead.  This
-  // will be equally fast, and create less vector register pressure.
-  if (Indices.size() == 1 && !Subtarget.hasOptimizedSegmentLoadStore(Factor)) {
-    unsigned ScalarSizeInBytes = DL.getTypeStoreSize(VTy->getElementType());
-    Value *Stride = ConstantInt::get(XLenTy, Factor * ScalarSizeInBytes);
-    Value *Offset = ConstantInt::get(XLenTy, Indices[0] * ScalarSizeInBytes);
-    Value *BasePtr = Builder.CreatePtrAdd(Ptr, Offset);
-    // For rv64, need to truncate i64 to i32 to match signature.  As VL is at most
-    // the number of active lanes (which is bounded by i32) this is safe.
-    VL = Builder.CreateTrunc(VL, Builder.getInt32Ty());
-
-    CallInst *CI =
-        Builder.CreateIntrinsic(Intrinsic::experimental_vp_strided_load,
-                                {VTy, BasePtr->getType(), Stride->getType()},
-                                {BasePtr, Stride, Mask, VL});
-    Alignment = commonAlignment(Alignment, Indices[0] * ScalarSizeInBytes);
-    CI->addParamAttr(0,
-                     Attribute::getWithAlignment(CI->getContext(), Alignment));
-    Shuffles[0]->replaceAllUsesWith(CI);
-    return true;
-  };
-
   CallInst *VlsegN = Builder.CreateIntrinsic(
       FixedVlsegIntrIds[Factor - 2], {VTy, PtrTy, XLenTy}, {Ptr, Mask, VL});
 

llvm/test/CodeGen/RISCV/rvv/pr141907.ll

@@ -9,27 +9,29 @@ define void @pr141907(ptr %0) nounwind {
 ; CHECK-NEXT:    slli a1, a1, 2
 ; CHECK-NEXT:    sub sp, sp, a1
 ; CHECK-NEXT:    vsetivli zero, 0, e32, m1, ta, ma
-; CHECK-NEXT:    vmv.v.i v9, 0
+; CHECK-NEXT:    vmv.v.i v8, 0
 ; CHECK-NEXT:    vmclr.m v0
 ; CHECK-NEXT:    li a1, 0
-; CHECK-NEXT:    vsetvli a3, zero, e16, mf2, ta, ma
-; CHECK-NEXT:    vmv.v.i v12, 0
+; CHECK-NEXT:    vsetvli a5, zero, e16, mf2, ta, ma
+; CHECK-NEXT:    vmv.v.i v10, 0
 ; CHECK-NEXT:    addi a2, sp, 16
+; CHECK-NEXT:    addi a3, sp, 20
+; CHECK-NEXT:    li a4, 12
 ; CHECK-NEXT:  .LBB0_1: # %vector.body
 ; CHECK-NEXT:    # =>This Inner Loop Header: Depth=1
 ; CHECK-NEXT:    vs4r.v v8, (a2)
 ; CHECK-NEXT:    vsetvli a1, a1, e8, mf8, ta, ma
 ; CHECK-NEXT:    vsetivli zero, 0, e16, mf2, ta, ma
-; CHECK-NEXT:    vnsrl.wi v11, v9, 0, v0.t
-; CHECK-NEXT:    vsetvli a3, zero, e32, m1, ta, ma
-; CHECK-NEXT:    vlseg3e32.v v8, (a2)
+; CHECK-NEXT:    vnsrl.wi v9, v8, 0, v0.t
+; CHECK-NEXT:    vsetvli a5, zero, e32, m1, ta, ma
+; CHECK-NEXT:    vlse32.v v8, (a3), a4
 ; CHECK-NEXT:    vsetivli zero, 0, e16, mf2, ta, ma
-; CHECK-NEXT:    vsseg2e16.v v11, (zero)
+; CHECK-NEXT:    vsseg2e16.v v9, (zero)
 ; CHECK-NEXT:    bnez a1, .LBB0_1
 ; CHECK-NEXT:  .LBB0_2: # %while.body5
 ; CHECK-NEXT:    # =>This Inner Loop Header: Depth=1
 ; CHECK-NEXT:    vsetivli zero, 1, e16, m1, ta, ma
-; CHECK-NEXT:    vse16.v v9, (a0)
+; CHECK-NEXT:    vse16.v v8, (a0)
 ; CHECK-NEXT:    j .LBB0_2
 entry:
   br label %vector.body

llvm/test/CodeGen/RISCV/rvv/vector-deinterleave-load.ll

@@ -407,8 +407,9 @@ define { <vscale x 8 x i8>, <vscale x 8 x i8>, <vscale x 8 x i8>, <vscale x 8 x
 define <vscale x 8 x i8> @vector_deinterleave_load_factor4_oneactive(ptr %p) {
 ; CHECK-LABEL: vector_deinterleave_load_factor4_oneactive:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a1, zero, e8, m1, ta, ma
-; CHECK-NEXT:    vlseg4e8.v v8, (a0)
+; CHECK-NEXT:    li a1, 4
+; CHECK-NEXT:    vsetvli a2, zero, e8, m1, ta, ma
+; CHECK-NEXT:    vlse8.v v8, (a0), a1
 ; CHECK-NEXT:    ret
   %vec = load <vscale x 32 x i8>, ptr %p
   %d0 = call { <vscale x 8 x i8>, <vscale x 8 x i8>, <vscale x 8 x i8>, <vscale x 8 x i8> } @llvm.vector.deinterleave4(<vscale x 32 x i8> %vec)
@@ -419,8 +420,10 @@ define <vscale x 8 x i8> @vector_deinterleave_load_factor4_oneactive(ptr %p) {
 define <vscale x 8 x i8> @vector_deinterleave_load_factor4_oneactive2(ptr %p) {
 ; CHECK-LABEL: vector_deinterleave_load_factor4_oneactive2:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a1, zero, e8, m1, ta, ma
-; CHECK-NEXT:    vlseg4e8.v v5, (a0)
+; CHECK-NEXT:    addi a0, a0, 3
+; CHECK-NEXT:    li a1, 4
+; CHECK-NEXT:    vsetvli a2, zero, e8, m1, ta, ma
+; CHECK-NEXT:    vlse8.v v8, (a0), a1
 ; CHECK-NEXT:    ret
   %vec = load <vscale x 32 x i8>, ptr %p
   %d0 = call { <vscale x 8 x i8>, <vscale x 8 x i8>, <vscale x 8 x i8>, <vscale x 8 x i8> } @llvm.vector.deinterleave4(<vscale x 32 x i8> %vec)

llvm/test/CodeGen/RISCV/rvv/vector-deinterleave.ll

@@ -3712,8 +3712,9 @@ define <vscale x 1 x float> @vector_deinterleave_nxv1f32_nxv8f32_oneactive(<vsca
 ; CHECK-NEXT:    sub sp, sp, a0
 ; CHECK-NEXT:    addi a0, sp, 16
 ; CHECK-NEXT:    vs4r.v v8, (a0)
-; CHECK-NEXT:    vsetvli a1, zero, e32, mf2, ta, ma
-; CHECK-NEXT:    vlseg8e32.v v8, (a0)
+; CHECK-NEXT:    li a1, 32
+; CHECK-NEXT:    vsetvli a2, zero, e32, mf2, ta, ma
+; CHECK-NEXT:    vlse32.v v8, (a0), a1
 ; CHECK-NEXT:    csrr a0, vlenb
 ; CHECK-NEXT:    slli a0, a0, 2
 ; CHECK-NEXT:    add sp, sp, a0
@@ -3732,9 +3733,11 @@ define <vscale x 1 x float> @vector_deinterleave_nxv1f32_nxv8f32_oneactive2(<vsc
 ; CHECK-NEXT:    slli a0, a0, 2
 ; CHECK-NEXT:    sub sp, sp, a0
 ; CHECK-NEXT:    addi a0, sp, 16
+; CHECK-NEXT:    addi a1, sp, 36
 ; CHECK-NEXT:    vs4r.v v8, (a0)
-; CHECK-NEXT:    vsetvli a1, zero, e32, mf2, ta, ma
-; CHECK-NEXT:    vlseg8e32.v v3, (a0)
+; CHECK-NEXT:    li a0, 32
+; CHECK-NEXT:    vsetvli a2, zero, e32, mf2, ta, ma
+; CHECK-NEXT:    vlse32.v v8, (a1), a0
 ; CHECK-NEXT:    csrr a0, vlenb
 ; CHECK-NEXT:    slli a0, a0, 2
 ; CHECK-NEXT:    add sp, sp, a0

llvm/test/CodeGen/RISCV/rvv/vp-vector-interleaved-access.ll

@@ -674,16 +674,20 @@ define <vscale x 2 x i32> @load_factor2_oneactive(ptr %ptr, i32 %evl) {
 define <vscale x 2 x i32> @load_factor5_oneactive(ptr %ptr, i32 %evl) {
 ; RV32-LABEL: load_factor5_oneactive:
 ; RV32:       # %bb.0:
+; RV32-NEXT:    addi a0, a0, 12
+; RV32-NEXT:    li a2, 20
 ; RV32-NEXT:    vsetvli zero, a1, e32, m1, ta, ma
-; RV32-NEXT:    vlseg5e32.v v5, (a0)
+; RV32-NEXT:    vlse32.v v8, (a0), a2
 ; RV32-NEXT:    ret
 ;
 ; RV64-LABEL: load_factor5_oneactive:
 ; RV64:       # %bb.0:
 ; RV64-NEXT:    slli a1, a1, 32
+; RV64-NEXT:    addi a0, a0, 12
 ; RV64-NEXT:    srli a1, a1, 32
+; RV64-NEXT:    li a2, 20
 ; RV64-NEXT:    vsetvli zero, a1, e32, m1, ta, ma
-; RV64-NEXT:    vlseg5e32.v v5, (a0)
+; RV64-NEXT:    vlse32.v v8, (a0), a2
 ; RV64-NEXT:    ret
   %rvl = mul nuw i32 %evl, 5
   %wide.masked.load = call <vscale x 10 x i32> @llvm.vp.load(ptr %ptr, <vscale x 10 x i1> splat (i1 true), i32 %rvl)