[RISCV] Add side-effect-free vsetvli intrinsics

[frasercrmck]

Summary

This patch introduces new intrinsics that enable the use of vsetvli in
contexts where only the returned vector length is of interest. The
pre-existing intrinsics are marked with side-effects, which prevents
even trivial optimizations on/across them.

These intrinsics are intended to be used in situations where the vector
length is fed in turn to RVV intrinsics or to vector-predication
intrinsics during loop vectorization, for example. Those codegen paths
ensure that instructions are generated with their own implicit vsetvli,
so the vector length and vtype can be relied upon to be correct.

No corresponding C builtins are planned at this stage, though that is a
possibility for the future if the need arises.

[frasercrmck]

I apologise that opt is so uninspired a name. I'll blame it on the fact it's Friday. I'm very open to alternatives!

[craig.topper]

LGTM

[kito-cheng]

I would prefer use that to replace existing vsetvlimax and vsetvli once we verified that, I believe vsetvl and vsetvlmax intrinsic should be no side effect on C/C++ language level.

[frasercrmck]

In D117910#3266105, @kito-cheng wrote:

I would prefer use that to replace existing vsetvlimax and vsetvli once we verified that, I believe vsetvl and vsetvlmax intrinsic should be no side effect on C/C++ language level.

Yes I would have preferred that too. However I spoke with Craig offline and he told me that there are users which have now come to depend on the C/C++ vsetvli/vsetvlimax builtins having side effects. So I fear that ship may have sailed? I'd definitely be up for discussing a transition path but we'd need to assess the implications and any transition path.

[jrtc27]

In D117910#3266144, @frasercrmck wrote:

In D117910#3266105, @kito-cheng wrote:

I would prefer use that to replace existing vsetvlimax and vsetvli once we verified that, I believe vsetvl and vsetvlmax intrinsic should be no side effect on C/C++ language level.

Yes I would have preferred that too. However I spoke with Craig offline and he told me that there are users which have now come to depend on the C/C++ vsetvli/vsetvlimax builtins having side effects. So I fear that ship may have sailed? I'd definitely be up for discussing a transition path but we'd need to assess the implications and any transition path.

The extensions were experimental until extremely recently (maybe V still is, even?); as someone not hugely involved in the vector work, if non-side-effecting intrinsics are significantly better than side-effecting ones then you should prioritise getting it right. The fallout is unfortunate but will be short-term, and we've never guaranteed any kind of stability for experimental extensions. Otherwise you'll have much more pain in the long run dealing with this legacy baggage.

[craig.topper]

In D117910#3266172, @jrtc27 wrote:

In D117910#3266144, @frasercrmck wrote:

In D117910#3266105, @kito-cheng wrote:

I would prefer use that to replace existing vsetvlimax and vsetvli once we verified that, I believe vsetvl and vsetvlmax intrinsic should be no side effect on C/C++ language level.

Yes I would have preferred that too. However I spoke with Craig offline and he told me that there are users which have now come to depend on the C/C++ vsetvli/vsetvlimax builtins having side effects. So I fear that ship may have sailed? I'd definitely be up for discussing a transition path but we'd need to assess the implications and any transition path.

The extensions were experimental until extremely recently (maybe V still is, even?); as someone not hugely involved in the vector work, if non-side-effecting intrinsics are significantly better than side-effecting ones then you should prioritise getting it right. The fallout is unfortunate but will be short-term, and we've never guaranteed any kind of stability for experimental extensions. Otherwise you'll have much more pain in the long run dealing with this legacy baggage.

I have users who have used the side effecting property to prevent hoisting of V instructions like vmv.v.i by LICM when the VL is a constant. Because of where our VSETVLI insertion step happens in the pipeline, the register allocator is unable to rematerialize the vmv.v.i back to its uses to avoid copies. The VL physical register dependency prevents it. This is bad for register pressure, especially at larger LMUL values. By passing the constant VL to a sideeffecting vsetvli instruction and then passing the result to the vmv.v.i the hoisting is suppressed.

llvm/include/llvm/IR/IntrinsicsRISCV.td

def int_riscv_vsetvli   : Intrinsic<[llvm_anyint_ty],
                                        ImmArg<ArgIndex<2>>]>;
   def int_riscv_vsetvlimax : Intrinsic<[llvm_anyint_ty],
                             /* VSEW */ [LLVMMatchType<0>,
                             /* VLMUL */ LLVMMatchType<0>],
                                       [IntrNoMem, IntrHasSideEffects,
                                        ImmArg<ArgIndex<0>>,
                                        ImmArg<ArgIndex<1>>]>;
 
+  // Versions without side effects: better optimizable and usable if only the
+  // returned vector length is important.
+  def int_riscv_vsetvli_opt   : Intrinsic<[llvm_anyint_ty],
+                               /* AVL */  [LLVMMatchType<0>,
+                               /* VSEW */  LLVMMatchType<0>,
+                               /* VLMUL */ LLVMMatchType<0>],
+                                          [IntrNoMem,
+                                           ImmArg<ArgIndex<1>>,
+                                           ImmArg<ArgIndex<2>>]>;
+  def int_riscv_vsetvlimax_opt : Intrinsic<[llvm_anyint_ty],
+                                /* VSEW */ [LLVMMatchType<0>,
+                                /* VLMUL */ LLVMMatchType<0>],
+                                          [IntrNoMem,
+                                           ImmArg<ArgIndex<0>>,
+                                           ImmArg<ArgIndex<1>>]>;
+
   // For unit stride load
   // Input: (pointer, vl)
   class RISCVUSLoad
         : Intrinsic<[llvm_anyvector_ty],
                     [LLVMPointerType<LLVMMatchType<0>>,
                      llvm_anyint_ty],
                     [NoCapture<ArgIndex<0>>, IntrReadMem]>, RISCVVIntrinsic {
     let VLOperand = 1;

llvm/lib/Target/RISCV/RISCVISelDAGToDAG.h

void addVectorLoadStoreOperands(SDNode *Node, unsigned SEWImm,
                                   bool IsLoad = false, MVT *IndexVT = nullptr);
 
   void selectVLSEG(SDNode *Node, bool IsMasked, bool IsStrided);
   void selectVLSEGFF(SDNode *Node, bool IsMasked);
   void selectVLXSEG(SDNode *Node, bool IsMasked, bool IsOrdered);
   void selectVSSEG(SDNode *Node, bool IsMasked, bool IsStrided);
   void selectVSXSEG(SDNode *Node, bool IsMasked, bool IsOrdered);
 
+  void selectVSETVLI(SDNode *Node);
+
   // Return the RISC-V condition code that matches the given DAG integer
   // condition code. The CondCode must be one of those supported by the RISC-V
   // ISA (see translateSetCCForBranch).
   static RISCVCC::CondCode getRISCVCCForIntCC(ISD::CondCode CC) {
     switch (CC) {
     default:
       llvm_unreachable("Unsupported CondCode");
     case ISD::SETEQ:

llvm/lib/Target/RISCV/RISCVISelDAGToDAG.cpp

MachineSDNode *Store =
       CurDAG->getMachineNode(P->Pseudo, DL, Node->getValueType(0), Operands);
 
   if (auto *MemOp = dyn_cast<MemSDNode>(Node))
     CurDAG->setNodeMemRefs(Store, {MemOp->getMemOperand()});
 
   ReplaceNode(Node, Store);
 }
 
+void RISCVDAGToDAGISel::selectVSETVLI(SDNode *Node) {
+  if (!Subtarget->hasVInstructions())
+    return;
+
+  assert((Node->getOpcode() == ISD::INTRINSIC_W_CHAIN ||
+          Node->getOpcode() == ISD::INTRINSIC_WO_CHAIN) &&
+         "Unexpected opcode");
+
+  SDLoc DL(Node);
+  MVT XLenVT = Subtarget->getXLenVT();
+
+  bool HasChain = Node->getOpcode() == ISD::INTRINSIC_W_CHAIN;
+  unsigned IntNoOffset = HasChain ? 1 : 0;
+  unsigned IntNo = Node->getConstantOperandVal(IntNoOffset);
+
+  assert((IntNo == Intrinsic::riscv_vsetvli ||
+          IntNo == Intrinsic::riscv_vsetvlimax ||
+          IntNo == Intrinsic::riscv_vsetvli_opt ||
+          IntNo == Intrinsic::riscv_vsetvlimax_opt) &&
+         "Unexpected vsetvli intrinsic");
+
+  bool VLMax = IntNo == Intrinsic::riscv_vsetvlimax ||
+               IntNo == Intrinsic::riscv_vsetvlimax_opt;
+  unsigned Offset = IntNoOffset + (VLMax ? 1 : 2);
+
+  assert(Node->getNumOperands() == Offset + 2 &&
+         "Unexpected number of operands");
+
+  unsigned SEW =
+      RISCVVType::decodeVSEW(Node->getConstantOperandVal(Offset) & 0x7);
+  RISCVII::VLMUL VLMul = static_cast<RISCVII::VLMUL>(
+      Node->getConstantOperandVal(Offset + 1) & 0x7);
+
+  unsigned VTypeI = RISCVVType::encodeVTYPE(VLMul, SEW, /*TailAgnostic*/ true,
+                                            /*MaskAgnostic*/ false);
+  SDValue VTypeIOp = CurDAG->getTargetConstant(VTypeI, DL, XLenVT);
+
+  SmallVector<EVT, 2> VTs = {XLenVT};
+  if (HasChain)
+    VTs.push_back(MVT::Other);
+
+  SDValue VLOperand;
+  unsigned Opcode = RISCV::PseudoVSETVLI;
+  if (VLMax) {
+    VLOperand = CurDAG->getRegister(RISCV::X0, XLenVT);
+    Opcode = RISCV::PseudoVSETVLIX0;
+  } else {
+    VLOperand = Node->getOperand(IntNoOffset + 1);
+
+    if (auto *C = dyn_cast<ConstantSDNode>(VLOperand)) {
+      uint64_t AVL = C->getZExtValue();
+      if (isUInt<5>(AVL)) {
+        SDValue VLImm = CurDAG->getTargetConstant(AVL, DL, XLenVT);
+        SmallVector<SDValue, 3> Ops = {VLImm, VTypeIOp};
+        if (HasChain)
+          Ops.push_back(Node->getOperand(0));
+        ReplaceNode(
+            Node, CurDAG->getMachineNode(RISCV::PseudoVSETIVLI, DL, VTs, Ops));
+        return;
+      }
+    }
+  }
+
+  SmallVector<SDValue, 3> Ops = {VLOperand, VTypeIOp};
+  if (HasChain)
+    Ops.push_back(Node->getOperand(0));
+
+  ReplaceNode(Node, CurDAG->getMachineNode(Opcode, DL, VTs, Ops));
+}
 
 void RISCVDAGToDAGISel::Select(SDNode *Node) {
   // If we have a custom node, we have already selected.
   if (Node->isMachineOpcode()) {
     LLVM_DEBUG(dbgs() << "== "; Node->dump(CurDAG); dbgs() << "\n");
     Node->setNodeId(-1);
     return;
   }
#undef CASE_VMXOR_VMANDN_VMAND_OPCODES
       SDValue Cmp = SDValue(
           CurDAG->getMachineNode(VMSLTMaskOpcode, DL, VT,
                                  {MaskedOff, Src1, Src2, V0, VL, SEW, Glue}),
           0);
       ReplaceNode(Node, CurDAG->getMachineNode(VMXOROpcode, DL, VT,
                                                {Cmp, Mask, VL, MaskSEW}));
       return;
     }
+    case Intrinsic::riscv_vsetvli_opt:
+    case Intrinsic::riscv_vsetvlimax_opt:
+      return selectVSETVLI(Node);
     }
     break;
   }
   case ISD::INTRINSIC_W_CHAIN: {
     unsigned IntNo = cast<ConstantSDNode>(Node->getOperand(1))->getZExtValue();
     switch (IntNo) {
       // By default we do not custom select any intrinsic.
     default:
       break;
-
     case Intrinsic::riscv_vsetvli:
-    case Intrinsic::riscv_vsetvlimax: {
-      if (!Subtarget->hasVInstructions())
-        break;
-
-      bool VLMax = IntNo == Intrinsic::riscv_vsetvlimax;
-      unsigned Offset = VLMax ? 2 : 3;
-
-      assert(Node->getNumOperands() == Offset + 2 &&
-             "Unexpected number of operands");
-
-      unsigned SEW =
-          RISCVVType::decodeVSEW(Node->getConstantOperandVal(Offset) & 0x7);
-      RISCVII::VLMUL VLMul = static_cast<RISCVII::VLMUL>(
-          Node->getConstantOperandVal(Offset + 1) & 0x7);
-
-      unsigned VTypeI = RISCVVType::encodeVTYPE(
-          VLMul, SEW, /*TailAgnostic*/ true, /*MaskAgnostic*/ false);
-      SDValue VTypeIOp = CurDAG->getTargetConstant(VTypeI, DL, XLenVT);
-
-      SDValue VLOperand;
-      unsigned Opcode = RISCV::PseudoVSETVLI;
-      if (VLMax) {
-        VLOperand = CurDAG->getRegister(RISCV::X0, XLenVT);
-        Opcode = RISCV::PseudoVSETVLIX0;
-      } else {
-        VLOperand = Node->getOperand(2);
-
-        if (auto *C = dyn_cast<ConstantSDNode>(VLOperand)) {
-          uint64_t AVL = C->getZExtValue();
-          if (isUInt<5>(AVL)) {
-            SDValue VLImm = CurDAG->getTargetConstant(AVL, DL, XLenVT);
-            ReplaceNode(
-                Node, CurDAG->getMachineNode(RISCV::PseudoVSETIVLI, DL, XLenVT,
-                                             MVT::Other, VLImm, VTypeIOp,
-                                             /* Chain */ Node->getOperand(0)));
-            return;
-          }
-        }
-      }
-
-      ReplaceNode(Node,
-                  CurDAG->getMachineNode(Opcode, DL, XLenVT,
-                                         MVT::Other, VLOperand, VTypeIOp,
-                                         /* Chain */ Node->getOperand(0)));
-      return;
-    }
+    case Intrinsic::riscv_vsetvlimax:
+      return selectVSETVLI(Node);
     case Intrinsic::riscv_vlseg2:
     case Intrinsic::riscv_vlseg3:
     case Intrinsic::riscv_vlseg4:
     case Intrinsic::riscv_vlseg5:
     case Intrinsic::riscv_vlseg6:
     case Intrinsic::riscv_vlseg7:
     case Intrinsic::riscv_vlseg8: {
       selectVLSEG(Node, /*IsMasked*/ false, /*IsStrided*/ false);

llvm/lib/Target/RISCV/RISCVISelLowering.cpp

case RISCVISD::GREVW: {
     break;
   }
   case RISCVISD::READ_VLENB:
     // We assume VLENB is at least 16 bytes.
     Known.Zero.setLowBits(4);
     // We assume VLENB is no more than 65536 / 8 bytes.
     Known.Zero.setBitsFrom(14);
     break;
-  case ISD::INTRINSIC_W_CHAIN: {
-    unsigned IntNo = Op.getConstantOperandVal(1);
+  case ISD::INTRINSIC_W_CHAIN:
+  case ISD::INTRINSIC_WO_CHAIN: {
+    unsigned IntNo =
+        Op.getConstantOperandVal(Opc == ISD::INTRINSIC_WO_CHAIN ? 0 : 1);
     switch (IntNo) {
     default:
       // We can't do anything for most intrinsics.
       break;
     case Intrinsic::riscv_vsetvli:
     case Intrinsic::riscv_vsetvlimax:
+    case Intrinsic::riscv_vsetvli_opt:
+    case Intrinsic::riscv_vsetvlimax_opt:
       // Assume that VL output is positive and would fit in an int32_t.
       // TODO: VLEN might be capped at 16 bits in a future V spec update.
       if (BitWidth >= 32)
         Known.Zero.setBitsFrom(31);
       break;
     }
     break;
   }

llvm/test/CodeGen/RISCV/rvv/rv32-vsetvli-intrinsics.ll

 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc -mtriple=riscv32 -mattr=+v -verify-machineinstrs < %s | FileCheck %s
 
 declare i32 @llvm.riscv.vsetvli.i32(i32, i32, i32)
 declare i32 @llvm.riscv.vsetvlimax.i32(i32, i32)
+declare i32 @llvm.riscv.vsetvli.opt.i32(i32, i32, i32)
+declare i32 @llvm.riscv.vsetvlimax.opt.i32(i32, i32)
 
 define void @test_vsetvli_e64mf8(i32 %avl) nounwind {
 ; CHECK-LABEL: test_vsetvli_e64mf8:
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    vsetvli a0, a0, e64, mf8, ta, mu
 ; CHECK-NEXT:    ret
   call i32 @llvm.riscv.vsetvli.i32(i32 %avl, i32 3, i32 5)
   ret void

 ; CHECK-LABEL: test_vsetvlimax_e64m8:
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    vsetvli a0, zero, e64, m8, ta, mu
 ; CHECK-NEXT:    ret
   call i32 @llvm.riscv.vsetvlimax.i32(i32 3, i32 3)
   ret void
 }
 
+define i32 @test_vsetvli_opt_e8m1(i32 %avl) nounwind {
+; CHECK-LABEL: test_vsetvli_opt_e8m1:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vsetvli a0, a0, e8, m1, ta, mu
+; CHECK-NEXT:    ret
+  %vl = call i32 @llvm.riscv.vsetvli.opt.i32(i32 %avl, i32 0, i32 0)
+  ret i32 %vl
+}
+
+; Check that we remove the intrinsic if it's unused.
+define void @test_vsetvli_opt_e8m1_nouse(i32 %avl) nounwind {
+; CHECK-LABEL: test_vsetvli_opt_e8m1_nouse:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    ret
+  call i32 @llvm.riscv.vsetvli.opt.i32(i32 %avl, i32 0, i32 0)
+  ret void
+}
+
+define i32 @test_vsetvli_opt_e16mf4(i32 %avl) nounwind {
+; CHECK-LABEL: test_vsetvli_opt_e16mf4:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vsetvli a0, a0, e16, mf4, ta, mu
+; CHECK-NEXT:    ret
+  %vl = call i32 @llvm.riscv.vsetvli.opt.i32(i32 %avl, i32 1, i32 6)
+  ret i32 %vl
+}
+
+define i32 @test_vsetvli_opt_e32mf8_zero_avl() nounwind {
+; CHECK-LABEL: test_vsetvli_opt_e32mf8_zero_avl:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vsetivli a0, 0, e16, mf4, ta, mu
+; CHECK-NEXT:    ret
+  %vl = call i32 @llvm.riscv.vsetvli.opt.i32(i32 0, i32 1, i32 6)
+  ret i32 %vl
+}
+
+define i32 @test_vsetvlimax_opt_e32m2() nounwind {
+; CHECK-LABEL: test_vsetvlimax_opt_e32m2:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vsetvli a0, zero, e32, m2, ta, mu
+; CHECK-NEXT:    ret
+  %vl = call i32 @llvm.riscv.vsetvlimax.opt.i32(i32 2, i32 1)
+  ret i32 %vl
+}
+
+define void @test_vsetvlimax_opt_e32m2_nouse() nounwind {
+; CHECK-LABEL: test_vsetvlimax_opt_e32m2_nouse:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    ret
+  call i32 @llvm.riscv.vsetvlimax.opt.i32(i32 2, i32 1)
+  ret void
+}
+
+define i32 @test_vsetvlimax_opt_e64m4() nounwind {
+; CHECK-LABEL: test_vsetvlimax_opt_e64m4:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vsetvli a0, zero, e64, m4, ta, mu
+; CHECK-NEXT:    ret
+  %vl = call i32 @llvm.riscv.vsetvlimax.opt.i32(i32 3, i32 2)
+  ret i32 %vl
+}
+
 declare <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i32(<vscale x 4 x i32>*, i32)
 
 ; Check that we remove the redundant vsetvli when followed by another operation
 define <vscale x 4 x i32> @redundant_vsetvli(i32 %avl, <vscale x 4 x i32>* %ptr) nounwind {
 ; CHECK-LABEL: redundant_vsetvli:
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    vsetvli a0, a0, e32, m2, ta, mu
 ; CHECK-NEXT:    vle32.v v8, (a1)

llvm/test/CodeGen/RISCV/rvv/rv64-vsetvli-intrinsics.ll

 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc -mtriple=riscv64 -mattr=+v -verify-machineinstrs < %s | FileCheck %s
 
 declare i64 @llvm.riscv.vsetvli.i64(i64, i64, i64)
 declare i64 @llvm.riscv.vsetvlimax.i64(i64, i64)
+declare i64 @llvm.riscv.vsetvli.opt.i64(i64, i64, i64)
+declare i64 @llvm.riscv.vsetvlimax.opt.i64(i64, i64)
 
 define void @test_vsetvli_e8m1(i64 %avl) nounwind {
 ; CHECK-LABEL: test_vsetvli_e8m1:
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    vsetvli a0, a0, e8, m1, ta, mu
 ; CHECK-NEXT:    ret
   call i64 @llvm.riscv.vsetvli.i64(i64 %avl, i64 0, i64 0)
   ret void

 ; CHECK-LABEL: test_vsetvlimax_e64m4:
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    vsetvli a0, zero, e64, m4, ta, mu
 ; CHECK-NEXT:    ret
   call i64 @llvm.riscv.vsetvlimax.i64(i64 3, i64 2)
   ret void
 }
 
+define i64 @test_vsetvli_opt_e8m1(i64 %avl) nounwind {
+; CHECK-LABEL: test_vsetvli_opt_e8m1:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vsetvli a0, a0, e8, m1, ta, mu
+; CHECK-NEXT:    ret
+  %vl = call i64 @llvm.riscv.vsetvli.opt.i64(i64 %avl, i64 0, i64 0)
+  ret i64 %vl
+}
+
+; Check that we remove the intrinsic if it's unused.
+define void @test_vsetvli_opt_e8m1_nouse(i64 %avl) nounwind {
+; CHECK-LABEL: test_vsetvli_opt_e8m1_nouse:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    ret
+  call i64 @llvm.riscv.vsetvli.opt.i64(i64 %avl, i64 0, i64 0)
+  ret void
+}
+
+define i64 @test_vsetvli_opt_e16mf4(i64 %avl) nounwind {
+; CHECK-LABEL: test_vsetvli_opt_e16mf4:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vsetvli a0, a0, e16, mf4, ta, mu
+; CHECK-NEXT:    ret
+  %vl = call i64 @llvm.riscv.vsetvli.opt.i64(i64 %avl, i64 1, i64 6)
+  ret i64 %vl
+}
+
+define i64 @test_vsetvli_opt_e32mf8_zero_avl() nounwind {
+; CHECK-LABEL: test_vsetvli_opt_e32mf8_zero_avl:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vsetivli a0, 0, e16, mf4, ta, mu
+; CHECK-NEXT:    ret
+  %vl = call i64 @llvm.riscv.vsetvli.opt.i64(i64 0, i64 1, i64 6)
+  ret i64 %vl
+}
+
+define i64 @test_vsetvlimax_opt_e32m2() nounwind {
+; CHECK-LABEL: test_vsetvlimax_opt_e32m2:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vsetvli a0, zero, e32, m2, ta, mu
+; CHECK-NEXT:    ret
+  %vl = call i64 @llvm.riscv.vsetvlimax.opt.i64(i64 2, i64 1)
+  ret i64 %vl
+}
+
+define void @test_vsetvlimax_opt_e32m2_nouse() nounwind {
+; CHECK-LABEL: test_vsetvlimax_opt_e32m2_nouse:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    ret
+  call i64 @llvm.riscv.vsetvlimax.opt.i64(i64 2, i64 1)
+  ret void
+}
+
+define i64 @test_vsetvlimax_opt_e64m4() nounwind {
+; CHECK-LABEL: test_vsetvlimax_opt_e64m4:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vsetvli a0, zero, e64, m4, ta, mu
+; CHECK-NEXT:    ret
+  %vl = call i64 @llvm.riscv.vsetvlimax.opt.i64(i64 3, i64 2)
+  ret i64 %vl
+}
+
 declare <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32>*, i64)
 
 ; Check that we remove the redundant vsetvli when followed by another operation
 define <vscale x 4 x i32> @redundant_vsetvli(i64 %avl, <vscale x 4 x i32>* %ptr) nounwind {
 ; CHECK-LABEL: redundant_vsetvli:
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    vsetvli a0, a0, e32, m2, ta, mu
 ; CHECK-NEXT:    vle32.v v8, (a1)