[llvm] [SelectionDAG] Fix SplitVecRes_VP_SPLICE for sub-byte (i1) element types (PR #200626)

Justin Lebar via llvm-commits llvm-commits at lists.llvm.org
Sat May 30 20:09:32 PDT 2026


https://github.com/jlebar updated https://github.com/llvm/llvm-project/pull/200626

>From 6b1598cadc4eb9043628a77ac920585fe1c0669f Mon Sep 17 00:00:00 2001
From: Justin Lebar <justin.lebar at gmail.com>
Date: Sat, 30 May 2026 14:00:26 -0700
Subject: [PATCH] [SelectionDAG] Fix SplitVecRes_VP_SPLICE for sub-byte (i1)
 element types

The stack splice addresses elements by byte offset (getVectorElementPointer)
and, for a negative offset, a stride of getScalarSizeInBits() / 8. Both break
for i1: getVectorElementPointer asserts, or the stride degenerates to 0.

Widen sub-byte elements to a byte integer (i1 -> i8), splice, then truncate
back, like SplitVecRes_VP_REVERSE.
---
 .../SelectionDAG/LegalizeVectorTypes.cpp      |  18 +-
 .../RISCV/rvv/vp-splice-mask-vectors.ll       | 353 ++++++++++++++++++
 2 files changed, 370 insertions(+), 1 deletion(-)

diff --git a/llvm/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp b/llvm/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp
index 02e25e5855577..5f885d73cd610 100644
--- a/llvm/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp
@@ -3565,6 +3565,18 @@ void DAGTypeLegalizer::SplitVecRes_VP_SPLICE(SDNode *N, SDValue &Lo,
   if (getTypeAction(EVL1.getValueType()) == TargetLowering::TypePromoteInteger)
     EVL1 = ZExtPromotedInteger(EVL1);
 
+  // The stack splice addresses elements by byte offset/stride, which breaks for
+  // a sub-byte element (e.g. i1): getVectorElementPointer asserts and the
+  // stride is 0. Widen to a byte integer, splice, then truncate back.
+  EVT OrigVT = VT;
+  if (!VT.getVectorElementType().isByteSized()) {
+    EVT WideEltVT = VT.getVectorElementType().changeTypeToInteger();
+    WideEltVT = WideEltVT.getRoundIntegerType(*DAG.getContext());
+    VT = VT.changeVectorElementType(*DAG.getContext(), WideEltVT);
+    V1 = DAG.getNode(ISD::ANY_EXTEND, DL, VT, V1);
+    V2 = DAG.getNode(ISD::ANY_EXTEND, DL, VT, V2);
+  }
+
   Align Alignment = DAG.getReducedAlign(VT, /*UseABI=*/false);
 
   EVT MemVT = EVT::getVectorVT(*DAG.getContext(), VT.getVectorElementType(),
@@ -3613,8 +3625,12 @@ void DAGTypeLegalizer::SplitVecRes_VP_SPLICE(SDNode *N, SDValue &Lo,
     Load = DAG.getLoadVP(VT, DL, StoreV2, StackPtr2, Mask, EVL2, LoadMMO);
   }
 
+  // Truncate back if we widened above.
+  if (OrigVT != VT)
+    Load = DAG.getNode(ISD::TRUNCATE, DL, OrigVT, Load);
+
   EVT LoVT, HiVT;
-  std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(VT);
+  std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(OrigVT);
   Lo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, LoVT, Load,
                    DAG.getVectorIdxConstant(0, DL));
   Hi =
diff --git a/llvm/test/CodeGen/RISCV/rvv/vp-splice-mask-vectors.ll b/llvm/test/CodeGen/RISCV/rvv/vp-splice-mask-vectors.ll
index 3215b4548243c..42edcbe8ee3da 100644
--- a/llvm/test/CodeGen/RISCV/rvv/vp-splice-mask-vectors.ll
+++ b/llvm/test/CodeGen/RISCV/rvv/vp-splice-mask-vectors.ll
@@ -918,4 +918,357 @@ define <vscale x 64 x i1> @test_vp_splice_nxv64i1_masked(<vscale x 64 x i1> %va,
   ret <vscale x 64 x i1> %v
 }
 
+; nxv128i1 is split via the generic SplitVecRes_VP_SPLICE stack round-trip. The
+; i1 elements must be widened to a byte type; otherwise getVectorElementPointer
+; asserts and the negative-offset stride is 0.
+define <vscale x 128 x i1> @test_vp_splice_nxv128i1(<vscale x 128 x i1> %va, <vscale x 128 x i1> %vb, i32 zeroext %evla, i32 zeroext %evlb) #0 {
+; NOVLDEP-LABEL: test_vp_splice_nxv128i1:
+; NOVLDEP:       # %bb.0:
+; NOVLDEP-NEXT:    csrr a2, vlenb
+; NOVLDEP-NEXT:    slli a4, a2, 4
+; NOVLDEP-NEXT:    addi a4, a4, -1
+; NOVLDEP-NEXT:    mv a5, a0
+; NOVLDEP-NEXT:    mv a3, a0
+; NOVLDEP-NEXT:    bltu a0, a4, .LBB21_2
+; NOVLDEP-NEXT:  # %bb.1:
+; NOVLDEP-NEXT:    mv a3, a4
+; NOVLDEP-NEXT:  .LBB21_2:
+; NOVLDEP-NEXT:    addi sp, sp, -80
+; NOVLDEP-NEXT:    .cfi_def_cfa_offset 80
+; NOVLDEP-NEXT:    sd ra, 72(sp) # 8-byte Folded Spill
+; NOVLDEP-NEXT:    sd s0, 64(sp) # 8-byte Folded Spill
+; NOVLDEP-NEXT:    .cfi_offset ra, -8
+; NOVLDEP-NEXT:    .cfi_offset s0, -16
+; NOVLDEP-NEXT:    addi s0, sp, 80
+; NOVLDEP-NEXT:    .cfi_def_cfa s0, 0
+; NOVLDEP-NEXT:    csrr a4, vlenb
+; NOVLDEP-NEXT:    slli a4, a4, 5
+; NOVLDEP-NEXT:    sub sp, sp, a4
+; NOVLDEP-NEXT:    andi sp, sp, -64
+; NOVLDEP-NEXT:    addi a4, sp, 64
+; NOVLDEP-NEXT:    vsetvli a6, zero, e8, m8, ta, ma
+; NOVLDEP-NEXT:    vmv.v.i v16, 0
+; NOVLDEP-NEXT:    slli a2, a2, 3
+; NOVLDEP-NEXT:    add a3, a4, a3
+; NOVLDEP-NEXT:    vmerge.vim v24, v16, 1, v0
+; NOVLDEP-NEXT:    bltu a5, a2, .LBB21_4
+; NOVLDEP-NEXT:  # %bb.3:
+; NOVLDEP-NEXT:    mv a5, a2
+; NOVLDEP-NEXT:  .LBB21_4:
+; NOVLDEP-NEXT:    vmv1r.v v0, v8
+; NOVLDEP-NEXT:    vsetvli zero, a5, e8, m8, ta, ma
+; NOVLDEP-NEXT:    vse8.v v24, (a4)
+; NOVLDEP-NEXT:    sub a5, a0, a2
+; NOVLDEP-NEXT:    add a4, a4, a2
+; NOVLDEP-NEXT:    sltu a0, a0, a5
+; NOVLDEP-NEXT:    addi a0, a0, -1
+; NOVLDEP-NEXT:    and a0, a0, a5
+; NOVLDEP-NEXT:    sub a5, a1, a2
+; NOVLDEP-NEXT:    vsetvli zero, a0, e8, m8, ta, ma
+; NOVLDEP-NEXT:    vmerge.vim v24, v16, 1, v0
+; NOVLDEP-NEXT:    vse8.v v24, (a4)
+; NOVLDEP-NEXT:    add a4, a3, a2
+; NOVLDEP-NEXT:    sltu a0, a1, a5
+; NOVLDEP-NEXT:    addi a0, a0, -1
+; NOVLDEP-NEXT:    and a0, a0, a5
+; NOVLDEP-NEXT:    vmv1r.v v0, v10
+; NOVLDEP-NEXT:    vsetvli zero, a0, e8, m8, ta, ma
+; NOVLDEP-NEXT:    vmerge.vim v24, v16, 1, v0
+; NOVLDEP-NEXT:    vse8.v v24, (a4)
+; NOVLDEP-NEXT:    vmv1r.v v0, v9
+; NOVLDEP-NEXT:    vsetvli a4, zero, e8, m8, ta, ma
+; NOVLDEP-NEXT:    vmerge.vim v8, v16, 1, v0
+; NOVLDEP-NEXT:    bltu a1, a2, .LBB21_6
+; NOVLDEP-NEXT:  # %bb.5:
+; NOVLDEP-NEXT:    mv a1, a2
+; NOVLDEP-NEXT:  .LBB21_6:
+; NOVLDEP-NEXT:    vsetvli zero, a1, e8, m8, ta, ma
+; NOVLDEP-NEXT:    vse8.v v8, (a3)
+; NOVLDEP-NEXT:    addi a3, sp, 69
+; NOVLDEP-NEXT:    add a2, a3, a2
+; NOVLDEP-NEXT:    vsetvli zero, a0, e8, m8, ta, ma
+; NOVLDEP-NEXT:    vle8.v v8, (a2)
+; NOVLDEP-NEXT:    vsetvli zero, a1, e8, m8, ta, ma
+; NOVLDEP-NEXT:    vle8.v v16, (a3)
+; NOVLDEP-NEXT:    vsetvli a0, zero, e8, m8, ta, ma
+; NOVLDEP-NEXT:    vand.vi v24, v8, 1
+; NOVLDEP-NEXT:    vmsne.vi v8, v24, 0
+; NOVLDEP-NEXT:    vand.vi v16, v16, 1
+; NOVLDEP-NEXT:    vmsne.vi v0, v16, 0
+; NOVLDEP-NEXT:    addi sp, s0, -80
+; NOVLDEP-NEXT:    .cfi_def_cfa sp, 80
+; NOVLDEP-NEXT:    ld ra, 72(sp) # 8-byte Folded Reload
+; NOVLDEP-NEXT:    ld s0, 64(sp) # 8-byte Folded Reload
+; NOVLDEP-NEXT:    .cfi_restore ra
+; NOVLDEP-NEXT:    .cfi_restore s0
+; NOVLDEP-NEXT:    addi sp, sp, 80
+; NOVLDEP-NEXT:    .cfi_def_cfa_offset 0
+; NOVLDEP-NEXT:    ret
+;
+; VLDEP-LABEL: test_vp_splice_nxv128i1:
+; VLDEP:       # %bb.0:
+; VLDEP-NEXT:    csrr a2, vlenb
+; VLDEP-NEXT:    slli a4, a2, 4
+; VLDEP-NEXT:    addi a4, a4, -1
+; VLDEP-NEXT:    mv a5, a0
+; VLDEP-NEXT:    mv a3, a0
+; VLDEP-NEXT:    bltu a0, a4, .LBB21_2
+; VLDEP-NEXT:  # %bb.1:
+; VLDEP-NEXT:    mv a3, a4
+; VLDEP-NEXT:  .LBB21_2:
+; VLDEP-NEXT:    addi sp, sp, -80
+; VLDEP-NEXT:    .cfi_def_cfa_offset 80
+; VLDEP-NEXT:    sd ra, 72(sp) # 8-byte Folded Spill
+; VLDEP-NEXT:    sd s0, 64(sp) # 8-byte Folded Spill
+; VLDEP-NEXT:    .cfi_offset ra, -8
+; VLDEP-NEXT:    .cfi_offset s0, -16
+; VLDEP-NEXT:    addi s0, sp, 80
+; VLDEP-NEXT:    .cfi_def_cfa s0, 0
+; VLDEP-NEXT:    csrr a4, vlenb
+; VLDEP-NEXT:    slli a4, a4, 5
+; VLDEP-NEXT:    sub sp, sp, a4
+; VLDEP-NEXT:    andi sp, sp, -64
+; VLDEP-NEXT:    addi a4, sp, 64
+; VLDEP-NEXT:    vsetvli a6, zero, e8, m8, ta, ma
+; VLDEP-NEXT:    vmv.v.i v16, 0
+; VLDEP-NEXT:    slli a2, a2, 3
+; VLDEP-NEXT:    add a3, a4, a3
+; VLDEP-NEXT:    vmerge.vim v24, v16, 1, v0
+; VLDEP-NEXT:    bltu a5, a2, .LBB21_4
+; VLDEP-NEXT:  # %bb.3:
+; VLDEP-NEXT:    mv a5, a2
+; VLDEP-NEXT:  .LBB21_4:
+; VLDEP-NEXT:    vmv1r.v v0, v8
+; VLDEP-NEXT:    vsetvli zero, a5, e8, m8, ta, ma
+; VLDEP-NEXT:    vse8.v v24, (a4)
+; VLDEP-NEXT:    sub a5, a0, a2
+; VLDEP-NEXT:    add a4, a4, a2
+; VLDEP-NEXT:    sltu a0, a0, a5
+; VLDEP-NEXT:    addi a0, a0, -1
+; VLDEP-NEXT:    and a0, a0, a5
+; VLDEP-NEXT:    sub a5, a1, a2
+; VLDEP-NEXT:    vsetvli zero, a0, e8, m8, ta, ma
+; VLDEP-NEXT:    vmerge.vim v24, v16, 1, v0
+; VLDEP-NEXT:    vse8.v v24, (a4)
+; VLDEP-NEXT:    add a4, a3, a2
+; VLDEP-NEXT:    sltu a0, a1, a5
+; VLDEP-NEXT:    addi a0, a0, -1
+; VLDEP-NEXT:    and a0, a0, a5
+; VLDEP-NEXT:    vmv1r.v v0, v10
+; VLDEP-NEXT:    vsetvli zero, a0, e8, m8, ta, ma
+; VLDEP-NEXT:    vmerge.vim v24, v16, 1, v0
+; VLDEP-NEXT:    vse8.v v24, (a4)
+; VLDEP-NEXT:    vmv1r.v v0, v9
+; VLDEP-NEXT:    vsetvli a4, zero, e8, m8, ta, ma
+; VLDEP-NEXT:    vmerge.vim v8, v16, 1, v0
+; VLDEP-NEXT:    bltu a1, a2, .LBB21_6
+; VLDEP-NEXT:  # %bb.5:
+; VLDEP-NEXT:    mv a1, a2
+; VLDEP-NEXT:  .LBB21_6:
+; VLDEP-NEXT:    vsetvli zero, a1, e8, m8, ta, ma
+; VLDEP-NEXT:    vse8.v v8, (a3)
+; VLDEP-NEXT:    addi a3, sp, 69
+; VLDEP-NEXT:    add a2, a3, a2
+; VLDEP-NEXT:    vsetvli zero, a0, e8, m8, ta, ma
+; VLDEP-NEXT:    vle8.v v8, (a2)
+; VLDEP-NEXT:    vsetvli zero, a1, e8, m8, ta, ma
+; VLDEP-NEXT:    vle8.v v16, (a3)
+; VLDEP-NEXT:    vsetvli a0, zero, e8, m8, ta, ma
+; VLDEP-NEXT:    vand.vi v24, v8, 1
+; VLDEP-NEXT:    vmsne.vi v8, v24, 0
+; VLDEP-NEXT:    vand.vi v16, v16, 1
+; VLDEP-NEXT:    vmsne.vi v0, v16, 0
+; VLDEP-NEXT:    addi sp, s0, -80
+; VLDEP-NEXT:    .cfi_def_cfa sp, 80
+; VLDEP-NEXT:    ld ra, 72(sp) # 8-byte Folded Reload
+; VLDEP-NEXT:    ld s0, 64(sp) # 8-byte Folded Reload
+; VLDEP-NEXT:    .cfi_restore ra
+; VLDEP-NEXT:    .cfi_restore s0
+; VLDEP-NEXT:    addi sp, sp, 80
+; VLDEP-NEXT:    .cfi_def_cfa_offset 0
+; VLDEP-NEXT:    ret
+  %v = call <vscale x 128 x i1> @llvm.experimental.vp.splice.nxv128i1(<vscale x 128 x i1> %va, <vscale x 128 x i1> %vb, i32 5, <vscale x 128 x i1> splat (i1 1), i32 %evla, i32 %evlb)
+  ret <vscale x 128 x i1> %v
+}
+
+define <vscale x 128 x i1> @test_vp_splice_nxv128i1_negative_offset(<vscale x 128 x i1> %va, <vscale x 128 x i1> %vb, i32 zeroext %evla, i32 zeroext %evlb) #0 {
+; NOVLDEP-LABEL: test_vp_splice_nxv128i1_negative_offset:
+; NOVLDEP:       # %bb.0:
+; NOVLDEP-NEXT:    csrr a3, vlenb
+; NOVLDEP-NEXT:    slli a4, a3, 4
+; NOVLDEP-NEXT:    addi a4, a4, -1
+; NOVLDEP-NEXT:    mv a6, a0
+; NOVLDEP-NEXT:    mv a2, a0
+; NOVLDEP-NEXT:    bltu a0, a4, .LBB22_2
+; NOVLDEP-NEXT:  # %bb.1:
+; NOVLDEP-NEXT:    mv a2, a4
+; NOVLDEP-NEXT:  .LBB22_2:
+; NOVLDEP-NEXT:    addi sp, sp, -80
+; NOVLDEP-NEXT:    .cfi_def_cfa_offset 80
+; NOVLDEP-NEXT:    sd ra, 72(sp) # 8-byte Folded Spill
+; NOVLDEP-NEXT:    sd s0, 64(sp) # 8-byte Folded Spill
+; NOVLDEP-NEXT:    .cfi_offset ra, -8
+; NOVLDEP-NEXT:    .cfi_offset s0, -16
+; NOVLDEP-NEXT:    addi s0, sp, 80
+; NOVLDEP-NEXT:    .cfi_def_cfa s0, 0
+; NOVLDEP-NEXT:    csrr a4, vlenb
+; NOVLDEP-NEXT:    slli a4, a4, 5
+; NOVLDEP-NEXT:    sub sp, sp, a4
+; NOVLDEP-NEXT:    andi sp, sp, -64
+; NOVLDEP-NEXT:    addi a5, sp, 64
+; NOVLDEP-NEXT:    vsetvli a4, zero, e8, m8, ta, ma
+; NOVLDEP-NEXT:    vmv.v.i v16, 0
+; NOVLDEP-NEXT:    slli a3, a3, 3
+; NOVLDEP-NEXT:    add a4, a5, a2
+; NOVLDEP-NEXT:    vmerge.vim v24, v16, 1, v0
+; NOVLDEP-NEXT:    bltu a6, a3, .LBB22_4
+; NOVLDEP-NEXT:  # %bb.3:
+; NOVLDEP-NEXT:    mv a6, a3
+; NOVLDEP-NEXT:  .LBB22_4:
+; NOVLDEP-NEXT:    vmv1r.v v0, v8
+; NOVLDEP-NEXT:    vsetvli zero, a6, e8, m8, ta, ma
+; NOVLDEP-NEXT:    vse8.v v24, (a5)
+; NOVLDEP-NEXT:    sub a6, a0, a3
+; NOVLDEP-NEXT:    add a5, a5, a3
+; NOVLDEP-NEXT:    sltu a0, a0, a6
+; NOVLDEP-NEXT:    addi a0, a0, -1
+; NOVLDEP-NEXT:    and a0, a0, a6
+; NOVLDEP-NEXT:    sub a6, a1, a3
+; NOVLDEP-NEXT:    vsetvli zero, a0, e8, m8, ta, ma
+; NOVLDEP-NEXT:    vmerge.vim v24, v16, 1, v0
+; NOVLDEP-NEXT:    vse8.v v24, (a5)
+; NOVLDEP-NEXT:    add a5, a4, a3
+; NOVLDEP-NEXT:    sltu a0, a1, a6
+; NOVLDEP-NEXT:    addi a0, a0, -1
+; NOVLDEP-NEXT:    and a0, a0, a6
+; NOVLDEP-NEXT:    vmv1r.v v0, v10
+; NOVLDEP-NEXT:    vsetvli zero, a0, e8, m8, ta, ma
+; NOVLDEP-NEXT:    vmerge.vim v24, v16, 1, v0
+; NOVLDEP-NEXT:    vse8.v v24, (a5)
+; NOVLDEP-NEXT:    vmv1r.v v0, v9
+; NOVLDEP-NEXT:    vsetvli a5, zero, e8, m8, ta, ma
+; NOVLDEP-NEXT:    vmerge.vim v8, v16, 1, v0
+; NOVLDEP-NEXT:    bltu a1, a3, .LBB22_6
+; NOVLDEP-NEXT:  # %bb.5:
+; NOVLDEP-NEXT:    mv a1, a3
+; NOVLDEP-NEXT:  .LBB22_6:
+; NOVLDEP-NEXT:    li a5, 5
+; NOVLDEP-NEXT:    vsetvli zero, a1, e8, m8, ta, ma
+; NOVLDEP-NEXT:    vse8.v v8, (a4)
+; NOVLDEP-NEXT:    bltu a2, a5, .LBB22_8
+; NOVLDEP-NEXT:  # %bb.7:
+; NOVLDEP-NEXT:    li a2, 5
+; NOVLDEP-NEXT:  .LBB22_8:
+; NOVLDEP-NEXT:    sub a4, a4, a2
+; NOVLDEP-NEXT:    add a3, a4, a3
+; NOVLDEP-NEXT:    vsetvli zero, a0, e8, m8, ta, ma
+; NOVLDEP-NEXT:    vle8.v v8, (a3)
+; NOVLDEP-NEXT:    vsetvli zero, a1, e8, m8, ta, ma
+; NOVLDEP-NEXT:    vle8.v v16, (a4)
+; NOVLDEP-NEXT:    vsetvli a0, zero, e8, m8, ta, ma
+; NOVLDEP-NEXT:    vand.vi v24, v8, 1
+; NOVLDEP-NEXT:    vmsne.vi v8, v24, 0
+; NOVLDEP-NEXT:    vand.vi v16, v16, 1
+; NOVLDEP-NEXT:    vmsne.vi v0, v16, 0
+; NOVLDEP-NEXT:    addi sp, s0, -80
+; NOVLDEP-NEXT:    .cfi_def_cfa sp, 80
+; NOVLDEP-NEXT:    ld ra, 72(sp) # 8-byte Folded Reload
+; NOVLDEP-NEXT:    ld s0, 64(sp) # 8-byte Folded Reload
+; NOVLDEP-NEXT:    .cfi_restore ra
+; NOVLDEP-NEXT:    .cfi_restore s0
+; NOVLDEP-NEXT:    addi sp, sp, 80
+; NOVLDEP-NEXT:    .cfi_def_cfa_offset 0
+; NOVLDEP-NEXT:    ret
+;
+; VLDEP-LABEL: test_vp_splice_nxv128i1_negative_offset:
+; VLDEP:       # %bb.0:
+; VLDEP-NEXT:    csrr a3, vlenb
+; VLDEP-NEXT:    slli a4, a3, 4
+; VLDEP-NEXT:    addi a4, a4, -1
+; VLDEP-NEXT:    mv a6, a0
+; VLDEP-NEXT:    mv a2, a0
+; VLDEP-NEXT:    bltu a0, a4, .LBB22_2
+; VLDEP-NEXT:  # %bb.1:
+; VLDEP-NEXT:    mv a2, a4
+; VLDEP-NEXT:  .LBB22_2:
+; VLDEP-NEXT:    addi sp, sp, -80
+; VLDEP-NEXT:    .cfi_def_cfa_offset 80
+; VLDEP-NEXT:    sd ra, 72(sp) # 8-byte Folded Spill
+; VLDEP-NEXT:    sd s0, 64(sp) # 8-byte Folded Spill
+; VLDEP-NEXT:    .cfi_offset ra, -8
+; VLDEP-NEXT:    .cfi_offset s0, -16
+; VLDEP-NEXT:    addi s0, sp, 80
+; VLDEP-NEXT:    .cfi_def_cfa s0, 0
+; VLDEP-NEXT:    csrr a4, vlenb
+; VLDEP-NEXT:    slli a4, a4, 5
+; VLDEP-NEXT:    sub sp, sp, a4
+; VLDEP-NEXT:    andi sp, sp, -64
+; VLDEP-NEXT:    addi a5, sp, 64
+; VLDEP-NEXT:    vsetvli a4, zero, e8, m8, ta, ma
+; VLDEP-NEXT:    vmv.v.i v16, 0
+; VLDEP-NEXT:    slli a3, a3, 3
+; VLDEP-NEXT:    add a4, a5, a2
+; VLDEP-NEXT:    vmerge.vim v24, v16, 1, v0
+; VLDEP-NEXT:    bltu a6, a3, .LBB22_4
+; VLDEP-NEXT:  # %bb.3:
+; VLDEP-NEXT:    mv a6, a3
+; VLDEP-NEXT:  .LBB22_4:
+; VLDEP-NEXT:    vmv1r.v v0, v8
+; VLDEP-NEXT:    vsetvli zero, a6, e8, m8, ta, ma
+; VLDEP-NEXT:    vse8.v v24, (a5)
+; VLDEP-NEXT:    sub a6, a0, a3
+; VLDEP-NEXT:    add a5, a5, a3
+; VLDEP-NEXT:    sltu a0, a0, a6
+; VLDEP-NEXT:    addi a0, a0, -1
+; VLDEP-NEXT:    and a0, a0, a6
+; VLDEP-NEXT:    sub a6, a1, a3
+; VLDEP-NEXT:    vsetvli zero, a0, e8, m8, ta, ma
+; VLDEP-NEXT:    vmerge.vim v24, v16, 1, v0
+; VLDEP-NEXT:    vse8.v v24, (a5)
+; VLDEP-NEXT:    add a5, a4, a3
+; VLDEP-NEXT:    sltu a0, a1, a6
+; VLDEP-NEXT:    addi a0, a0, -1
+; VLDEP-NEXT:    and a0, a0, a6
+; VLDEP-NEXT:    vmv1r.v v0, v10
+; VLDEP-NEXT:    vsetvli zero, a0, e8, m8, ta, ma
+; VLDEP-NEXT:    vmerge.vim v24, v16, 1, v0
+; VLDEP-NEXT:    vse8.v v24, (a5)
+; VLDEP-NEXT:    vmv1r.v v0, v9
+; VLDEP-NEXT:    vsetvli a5, zero, e8, m8, ta, ma
+; VLDEP-NEXT:    vmerge.vim v8, v16, 1, v0
+; VLDEP-NEXT:    bltu a1, a3, .LBB22_6
+; VLDEP-NEXT:  # %bb.5:
+; VLDEP-NEXT:    mv a1, a3
+; VLDEP-NEXT:  .LBB22_6:
+; VLDEP-NEXT:    li a5, 5
+; VLDEP-NEXT:    vsetvli zero, a1, e8, m8, ta, ma
+; VLDEP-NEXT:    vse8.v v8, (a4)
+; VLDEP-NEXT:    bltu a2, a5, .LBB22_8
+; VLDEP-NEXT:  # %bb.7:
+; VLDEP-NEXT:    li a2, 5
+; VLDEP-NEXT:  .LBB22_8:
+; VLDEP-NEXT:    sub a4, a4, a2
+; VLDEP-NEXT:    add a3, a4, a3
+; VLDEP-NEXT:    vsetvli zero, a0, e8, m8, ta, ma
+; VLDEP-NEXT:    vle8.v v8, (a3)
+; VLDEP-NEXT:    vsetvli zero, a1, e8, m8, ta, ma
+; VLDEP-NEXT:    vle8.v v16, (a4)
+; VLDEP-NEXT:    vsetvli a0, zero, e8, m8, ta, ma
+; VLDEP-NEXT:    vand.vi v24, v8, 1
+; VLDEP-NEXT:    vmsne.vi v8, v24, 0
+; VLDEP-NEXT:    vand.vi v16, v16, 1
+; VLDEP-NEXT:    vmsne.vi v0, v16, 0
+; VLDEP-NEXT:    addi sp, s0, -80
+; VLDEP-NEXT:    .cfi_def_cfa sp, 80
+; VLDEP-NEXT:    ld ra, 72(sp) # 8-byte Folded Reload
+; VLDEP-NEXT:    ld s0, 64(sp) # 8-byte Folded Reload
+; VLDEP-NEXT:    .cfi_restore ra
+; VLDEP-NEXT:    .cfi_restore s0
+; VLDEP-NEXT:    addi sp, sp, 80
+; VLDEP-NEXT:    .cfi_def_cfa_offset 0
+; VLDEP-NEXT:    ret
+  %v = call <vscale x 128 x i1> @llvm.experimental.vp.splice.nxv128i1(<vscale x 128 x i1> %va, <vscale x 128 x i1> %vb, i32 -5, <vscale x 128 x i1> splat (i1 1), i32 %evla, i32 %evlb)
+  ret <vscale x 128 x i1> %v
+}
+
 attributes #0 = { vscale_range(2,0) }



More information about the llvm-commits mailing list