[ARM] Fold select_cc(vecreduce_[u|s][min|max], x) into VMINV or VMAXV

[samtebbs]

Summary

This folds a select_cc and (select(set_cc) of a max or min vector reduction with a scalar value into a VMAXV or VMINV.

[dmgreen]

llvm/lib/Target/ARM/ARMISelLowering.cpp

5189

This could do with a few extra tests to make sure it's a min/max. LHS==TrueVal and RHS==FalseVal for example.

Plus it's worth checking if commuted min/max work too. min(vecreduce.min, x) or min(x, vecreduce.min)

[samtebbs]

Support commuted operands and add tests to cover them.

[samtebbs]

llvm/lib/Target/ARM/ARMISelLowering.cpp

5189

Thanks for pointing that out, done now.

[dmgreen]

llvm/lib/Target/ARM/ARMISelLowering.cpp

5188

It's might be worth pulling this out into a separate function, similar to isLowerSaturatingConditional or something like PerformSplittingToNarrowingStores where we return the new SDValue.

5220–5222

This could use std::swap?

5231–5233

I think we might need to check these things, otherwise we might be matching things incorrectly.

I believe the exact semantics of a vminv.s8 are that it reads the bottom 8 bits of Rn to do the final scalar min. So a 32bit min could actually produce a different value. We should make sure we are getting this correct too.

[samtebbs]

Fold using the LHS and check for LHS and RHS opcodes explicitly.

llvm/lib/Target/ARM/ARMISelLowering.cpp

5188

That's much cleaner! Done.

5220–5222

Ah I didn't know that existed. Done.

5231–5233

Thanks. I've added explicit checks for the LHS opcode and have made the lowering use the LHS instead of the TrueVal as the scalar so that we know the top 24 bits aren't set.

[dmgreen]

llvm/lib/Target/ARM/ARMISelLowering.cpp

5168

else return false;
Then it doesn't need the extra indenting.

5186

This looks like it need to be a bit stricter still. The AND needs to go with a umin/umax, and the sign extend with smin/smax (they shouldn't be the other way around). The AND needs a mask of the right size (it'll be 255 for i8 for example, which is the same as a "zero extend inreg"). The sign extend will have a type as the second argument, which should be the same as the vecreduce's scalar type.

Some of this might be difficult to get to come up in practice, but we should make sure it won't be subtly wrong and cause bugs.

[samtebbs]

Be a bit more strict.

[samtebbs]

llvm/lib/Target/ARM/ARMISelLowering.cpp

5168

Thumbs up

5186

I thought about those two but hoped that the IR type system would make those situations impossible. Will add these, thanks.

[samtebbs]

Allow inverted condition codes

[samtebbs]

Add truncate_sext test

[samtebbs]

Add type promotion checks and two new tests

[dmgreen]

llvm/lib/Target/ARM/ARMISelLowering.cpp

5143

Please clang-format, and possible just make it a static function?

5195

-> FalseVal->getOperand(0).getValueType().getVectorElementType();

5201

Hmm. This is getting complicated.

I would suggest changing this to something more like:

if (Type == i32)
  return Selected == Compared;
else
  // Check the opcodes are and/signexted/assertzext/assertsext.

But I don't think that will work very well. For i8/i16 I think we might actually need to be matching:

    t7: i32 = AssertZext t5, ValueType:ch:i8
      t17: i32 = vecreduce_umin t3
    t25: i32 = and t17, Constant:i32<255>
  t27: i32 = select_cc t25, t7, t17, t7, setult:ch
t21: i32 = and t27, Constant:i32<255>

Including that root "and" (which is the same as the two TrueVal/FalseVal sides of the select_cc being extended). But one loop I tried had no visible way I could see to prove the value needed to be a i16. It might be easier to do this from a combine, not during lowering. So before we have legalized the types.

What do you think about getting i32 working first with this patch, and doing i8/i16 as a followup? That way we can get all the boilerplate out of the way, and it's just this code that produces the ARMISD::VMINVs that we need to work further on.

[samtebbs]

Perform the folding in a DAG combination rather than lowering.

[dmgreen]

Yeah, nice. This is a lot easier to follow.

llvm/lib/Target/ARM/ARMISelLowering.cpp

12150

I think it needs to check that operand is a Setcc too, to be sure it's not something strange.

12223

The top bits of this are not read by the instruction. Can we change it to an ANY_EXTEND?

12224

I think the last operand isn't needed here.

16092

Move the hasMVEIntegerOps check into the function? So it can look like the rest.

[samtebbs]

llvm/lib/Target/ARM/ARMISelLowering.cpp

12150

Good idea.

12223

Sure.

12224

You are indeed correct.

[dmgreen]

One last thing I thought of. Because this is pre-lowering we have to be careful with illegal types.

Can you check the types are the ones we expect them to be (v16i8/v8i16/v4i32). And add some tests for v2i64 (which should be scalarized).

llvm/lib/Target/ARM/ARMISelLowering.h

758

This is no longer used?

[samtebbs]

In D87836#2308677, @dmgreen wrote:

One last thing I thought of. Because this is pre-lowering we have to be careful with illegal types.

Can you check the types are the ones we expect them to be (v16i8/v8i16/v4i32). And add some tests for v2i64 (which should be scalarized).

Thanks for noticing this. I added tests for v2i64 and it failed an assertion so I have added checks for the valid vector types.

[samtebbs]

Add check for the valid vector types.

[dmgreen]

Thanks. LGTM

llvm/lib/Target/ARM/ARMISelLowering.cpp

12211

-> getOperand(0).getValueType()

[samtebbs]

In D87836#2311422, @dmgreen wrote:

Thanks. LGTM

Thanks!

[dmajor]

This caused a failure on the bots: http://lab.llvm.org:8011/builders/clang-x86_64-debian-fast/builds/38144

[aemerson]

In D87836#2312239, @dmajor wrote:

This caused a failure on the bots: http://lab.llvm.org:8011/builders/clang-x86_64-debian-fast/builds/38144

Yep, I just reverted it on c9f5cdd4531

[samtebbs]

Hi Amara.

Apologies for this. The fix is just a matter of updating the two failing
tests. What is the process for getting the change back in?

Sam

[SjoerdMeijer]

In D87836#2314074, @samtebbs wrote:

Hi Amara.

Apologies for this. The fix is just a matter of updating the two failing
tests. What is the process for getting the change back in?

If it's just a trivial update of a test, you can just go ahead and recommit it. If you're in doubt, you can always put it for review again.

[samtebbs]

In D87836#2314079, @SjoerdMeijer wrote:

In D87836#2314074, @samtebbs wrote:

Hi Amara.

Apologies for this. The fix is just a matter of updating the two failing
tests. What is the process for getting the change back in?

If it's just a trivial update of a test, you can just go ahead and recommit it. If you're in doubt, you can always put it for review again.

Thanks Sjoerd, I have done so in 68e002e1819f1598fc6815226a353ad2f04cd509

llvm/lib/Target/ARM/ARMISelLowering.h

enum NodeType : unsigned {
       VMLALVs,      // Same as VMLAV but with i64, returning the low and
       VMLALVu,      //   high 32-bit halves of the sum
       VMLALVps,     // Same as VMLALV[su] with a v4i1 predicate mask
       VMLALVpu,
       VMLALVAs,     // Same as VMLALV but also add an input accumulator
       VMLALVAu,     //   provided as low and high halves
       VMLALVAps,    // Same as VMLALVA[su] with a v4i1 predicate mask
       VMLALVApu,
+      VMINVu,        // Find minimum unsigned value of a vector and register
+      VMINVs,        // Find minimum signed value of a vector and register
+      VMAXVu,        // Find maximum unsigned value of a vector and register
+      VMAXVs,        // Find maximum signed value of a vector and register
 
       SMULWB,       // Signed multiply word by half word, bottom
       SMULWT,       // Signed multiply word by half word, top
       UMLAL,        // 64bit Unsigned Accumulate Multiply
       SMLAL,        // 64bit Signed Accumulate Multiply
       UMAAL,        // 64-bit Unsigned Accumulate Accumulate Multiply
       SMLALBB,      // 64-bit signed accumulate multiply bottom, bottom 16
       SMLALBT,      // 64-bit signed accumulate multiply bottom, top 16
SDValue LowerToTLSExecModels(GlobalAddressSDNode *GA,
                                  TLSModel::Model model) const;
     SDValue LowerGlobalTLSAddressDarwin(SDValue Op, SelectionDAG &DAG) const;
     SDValue LowerGlobalTLSAddressWindows(SDValue Op, SelectionDAG &DAG) const;
     SDValue LowerGLOBAL_OFFSET_TABLE(SDValue Op, SelectionDAG &DAG) const;
     SDValue LowerBR_JT(SDValue Op, SelectionDAG &DAG) const;
     SDValue LowerSignedALUO(SDValue Op, SelectionDAG &DAG) const;
     SDValue LowerUnsignedALUO(SDValue Op, SelectionDAG &DAG) const;
     SDValue LowerSELECT(SDValue Op, SelectionDAG &DAG) const;
     SDValue LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const;
     SDValue LowerBRCOND(SDValue Op, SelectionDAG &DAG) const;
     SDValue LowerBR_CC(SDValue Op, SelectionDAG &DAG) const;
     SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const;
     SDValue LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const;
     SDValue LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const;
     SDValue LowerShiftRightParts(SDValue Op, SelectionDAG &DAG) const;
     SDValue LowerShiftLeftParts(SDValue Op, SelectionDAG &DAG) const;
     SDValue LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) const;

llvm/lib/Target/ARM/ARMISelLowering.cpp

const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
   case ARMISD::VMLALVs:       return "ARMISD::VMLALVs";
   case ARMISD::VMLALVu:       return "ARMISD::VMLALVu";
   case ARMISD::VMLALVps:      return "ARMISD::VMLALVps";
   case ARMISD::VMLALVpu:      return "ARMISD::VMLALVpu";
   case ARMISD::VMLALVAs:      return "ARMISD::VMLALVAs";
   case ARMISD::VMLALVAu:      return "ARMISD::VMLALVAu";
   case ARMISD::VMLALVAps:     return "ARMISD::VMLALVAps";
   case ARMISD::VMLALVApu:     return "ARMISD::VMLALVApu";
+  case ARMISD::VMINVu:        return "ARMISD::VMINVu";
+  case ARMISD::VMINVs:        return "ARMISD::VMINVs";
+  case ARMISD::VMAXVu:        return "ARMISD::VMAXVu";
+  case ARMISD::VMAXVs:        return "ARMISD::VMAXVs";
   case ARMISD::UMAAL:         return "ARMISD::UMAAL";
   case ARMISD::UMLAL:         return "ARMISD::UMLAL";
   case ARMISD::SMLAL:         return "ARMISD::SMLAL";
   case ARMISD::SMLALBB:       return "ARMISD::SMLALBB";
   case ARMISD::SMLALBT:       return "ARMISD::SMLALBT";
   case ARMISD::SMLALTB:       return "ARMISD::SMLALTB";
   case ARMISD::SMLALTT:       return "ARMISD::SMLALTT";
   case ARMISD::SMULWB:        return "ARMISD::SMULWB";
if (VT == MVT::f32)
     return !Subtarget->hasVFP2Base();
   if (VT == MVT::f64)
     return !Subtarget->hasFP64();
   if (VT == MVT::f16)
     return !Subtarget->hasFullFP16();
   return false;
 }
 
 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
   EVT VT = Op.getValueType();
   SDLoc dl(Op);
 
   // Try to convert two saturating conditional selects into a single SSAT
   SDValue SatValue;
   uint64_t SatConstant;
   bool SatUSat;
   if (((!Subtarget->isThumb() && Subtarget->hasV6Ops()) || Subtarget->isThumb2()) &&
       isSaturatingConditional(Op, SatValue, SatConstant, SatUSat)) {
     if (SatUSat)
       return DAG.getNode(ARMISD::USAT, dl, VT, SatValue,
                          DAG.getConstant(countTrailingOnes(SatConstant), dl, VT));
     else
       return DAG.getNode(ARMISD::SSAT, dl, VT, SatValue,
                          DAG.getConstant(countTrailingOnes(SatConstant), dl, VT));
   }
 
   // Try to convert expressions of the form x < k ? k : x (and similar forms)
   // into more efficient bit operations, which is possible when k is 0 or -1
   // On ARM and Thumb-2 which have flexible operand 2 this will result in
   // single instructions. On Thumb the shift and the bit operation will be two
   // instructions.
   // Only allow this transformation on full-width (32-bit) operations
   SDValue LowerSatConstant;
   if (VT == MVT::i32 &&
       isLowerSaturatingConditional(Op, SatValue, LowerSatConstant)) {
     SDValue ShiftV = DAG.getNode(ISD::SRA, dl, VT, SatValue,
                                  DAG.getConstant(31, dl, VT));
     if (isNullConstant(LowerSatConstant)) {
       SDValue NotShiftV = DAG.getNode(ISD::XOR, dl, VT, ShiftV,
                                       DAG.getAllOnesConstant(dl, VT));
       return DAG.getNode(ISD::AND, dl, VT, SatValue, NotShiftV);
     } else if (isAllOnesConstant(LowerSatConstant))
       return DAG.getNode(ISD::OR, dl, VT, SatValue, ShiftV);
   }
 
   SDValue LHS = Op.getOperand(0);
   SDValue RHS = Op.getOperand(1);
   ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
   SDValue TrueVal = Op.getOperand(2);
   SDValue FalseVal = Op.getOperand(3);
   ConstantSDNode *CFVal = dyn_cast<ConstantSDNode>(FalseVal);
   ConstantSDNode *CTVal = dyn_cast<ConstantSDNode>(TrueVal);
 
+  if (Subtarget->hasMVEIntegerOps()) {
+    unsigned int Opcode = 0;
+    if ((TrueVal->getOpcode() == ISD::VECREDUCE_UMIN ||
+         FalseVal->getOpcode() == ISD::VECREDUCE_UMIN) &&
+        CC == ISD::SETULT)
+      Opcode = ARMISD::VMINVu;
+    else if ((TrueVal->getOpcode() == ISD::VECREDUCE_SMIN ||
+              FalseVal->getOpcode() == ISD::VECREDUCE_SMIN) &&
+             CC == ISD::SETLT)
+      Opcode = ARMISD::VMINVs;
+    else if ((TrueVal->getOpcode() == ISD::VECREDUCE_UMAX ||
+              FalseVal->getOpcode() == ISD::VECREDUCE_UMAX) &&
+             CC == ISD::SETUGT)
+      Opcode = ARMISD::VMAXVu;
+    else if ((TrueVal->getOpcode() == ISD::VECREDUCE_SMAX ||
+              FalseVal->getOpcode() == ISD::VECREDUCE_SMAX) &&
+             CC == ISD::SETGT)
+      Opcode = ARMISD::VMAXVs;
+
+    if (Opcode) {
+      // Copy to avoid corrupting the variables in the case that we don't lower
+      // to a vminv/vmaxv
+      SDValue Left = LHS;
+      SDValue Right = RHS;
+      SDValue Scalar = TrueVal;
+      SDValue Reduction = FalseVal;
+      // Normalise to the right hand side being the vector reduction
+      switch (Scalar->getOpcode()) {
+      case ISD::VECREDUCE_UMIN:
+      case ISD::VECREDUCE_SMIN:
+      case ISD::VECREDUCE_UMAX:
+      case ISD::VECREDUCE_SMAX:
+        SDValue tmp = Left;
+        Left = Right;
+        Right = tmp;
+
+        tmp = Scalar;
+        Scalar = Reduction;
+        Reduction = tmp;
+        break;
+      }
+
+      // Make sure that the reduction is on the right side of the comparison and
+      // the scalar on the left. Sometimes the left and/or right side wrap the
+      // scalar and/or reduction in another operation, like an AND or
+      // sign-extension, so check the first operand as well
+      if ((Left == Scalar || Left->getOperand(0) == Scalar) &&
+          (Right == Reduction || Right->getOperand(0) == Reduction)) {
+        return DAG.getNode(Opcode, dl, VT, Scalar, Reduction->getOperand(0));
+      }
+    }
+  }
+
   if (Subtarget->hasV8_1MMainlineOps() && CFVal && CTVal &&
       LHS.getValueType() == MVT::i32 && RHS.getValueType() == MVT::i32) {
     unsigned TVal = CTVal->getZExtValue();
     unsigned FVal = CFVal->getZExtValue();
     unsigned Opcode = 0;
 
     if (TVal == ~FVal) {
       Opcode = ARMISD::CSINV;
static SDValue PerformVSELECTCombine(SDNode *N,
                                      const ARMSubtarget *Subtarget) {
   // Transforms vselect(not(cond), lhs, rhs) into vselect(cond, rhs, lhs).
   //
   // We need to re-implement this optimization here as the implementation in the
   // Target-Independent DAGCombiner does not handle the kind of constant we make
   // (it calls isConstOrConstSplat with AllowTruncation set to false - and for
   // good reason, allowing truncation there would break other targets).
   //
   // Currently, this is only done for MVE, as it's the only target that benefits
   // from this transformation (e.g. VPNOT+VPSEL becomes a single VPSEL).
   if (!Subtarget->hasMVEIntegerOps())
     return SDValue();
 
   if (N->getOperand(0).getOpcode() != ISD::XOR)
     return SDValue();
   SDValue XOR = N->getOperand(0);
 
static SDValue PerformADDECombine(SDNode *N,
   if (DCI.isBeforeLegalize()) return SDValue();
 
   return AddCombineTo64bitUMAAL(N, DCI, Subtarget);
 }
 
 /// PerformADDCombineWithOperands - Try DAG combinations for an ADD with
 /// operands N0 and N1.  This is a helper for PerformADDCombine that is
 /// called with the default operands, and if that fails, with commuted
 /// operands.
 static SDValue PerformADDCombineWithOperands(SDNode *N, SDValue N0, SDValue N1,
                                           TargetLowering::DAGCombinerInfo &DCI,
                                           const ARMSubtarget *Subtarget){
   // Attempt to create vpadd for this add.
   if (SDValue Result = AddCombineToVPADD(N, N0, N1, DCI, Subtarget))
     return Result;
 
   // Attempt to create vpaddl for this add.
   if (SDValue Result = AddCombineVUZPToVPADDL(N, N0, N1, DCI, Subtarget))
     return Result;
   if (SDValue Result = AddCombineBUILD_VECTORToVPADDL(N, N0, N1, DCI,
                                                       Subtarget))
     return Result;
 
   // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
   if (N0.getNode()->hasOneUse())
     if (SDValue Result = combineSelectAndUse(N, N0, N1, DCI))
       return Result;
   return SDValue();
 }
 

 }
 
 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
                                              DAGCombinerInfo &DCI) const {
   switch (N->getOpcode()) {
   default: break;
   case ISD::VSELECT:    return PerformVSELECTCombine(N, DCI, Subtarget);
   case ISD::ABS:        return PerformABSCombine(N, DCI, Subtarget);
   case ARMISD::ADDE:    return PerformADDECombine(N, DCI, Subtarget);
   case ARMISD::UMLAL:   return PerformUMLALCombine(N, DCI.DAG, Subtarget);
   case ISD::ADD:        return PerformADDCombine(N, DCI, Subtarget);
   case ISD::SUB:        return PerformSUBCombine(N, DCI, Subtarget);
   case ISD::MUL:        return PerformMULCombine(N, DCI, Subtarget);
   case ISD::OR:         return PerformORCombine(N, DCI, Subtarget);
   case ISD::XOR:        return PerformXORCombine(N, DCI, Subtarget);
   case ISD::AND:        return PerformANDCombine(N, DCI, Subtarget);
   case ISD::BRCOND:

llvm/lib/Target/ARM/ARMInstrMVE.td

multiclass MVE_VMINMAXV_ty<string iname, bit isMin, string intrBaseName> {
   defm s8 : MVE_VMINMAXV_p<iname, 1, isMin, MVE_v16s8, intrBaseName>;
   defm s16: MVE_VMINMAXV_p<iname, 1, isMin, MVE_v8s16, intrBaseName>;
   defm s32: MVE_VMINMAXV_p<iname, 1, isMin, MVE_v4s32, intrBaseName>;
   defm u8 : MVE_VMINMAXV_p<iname, 1, isMin, MVE_v16u8, intrBaseName>;
   defm u16: MVE_VMINMAXV_p<iname, 1, isMin, MVE_v8u16, intrBaseName>;
   defm u32: MVE_VMINMAXV_p<iname, 1, isMin, MVE_v4u32, intrBaseName>;
 }
 
+def SDTVecReduceR : SDTypeProfile<1, 2, [   // Reduction of an integer and vector into an integer
+  SDTCisInt<0>, SDTCisInt<1>, SDTCisVec<2>
+]>;
+def ARMVMINVu       : SDNode<"ARMISD::VMINVu", SDTVecReduceR>;
+def ARMVMINVs       : SDNode<"ARMISD::VMINVs", SDTVecReduceR>;
+def ARMVMAXVu       : SDNode<"ARMISD::VMAXVu", SDTVecReduceR>;
+def ARMVMAXVs       : SDNode<"ARMISD::VMAXVs", SDTVecReduceR>;
+
 defm MVE_VMINV : MVE_VMINMAXV_ty<"vminv", 1, "int_arm_mve_minv">;
 defm MVE_VMAXV : MVE_VMINMAXV_ty<"vmaxv", 0, "int_arm_mve_maxv">;
 
 let Predicates = [HasMVEInt] in {
   def : Pat<(i32 (vecreduce_smax (v16i8 MQPR:$src))),
             (i32 (MVE_VMAXVs8 (t2MVNi (i32 127)), $src))>;
   def : Pat<(i32 (vecreduce_smax (v8i16 MQPR:$src))),
             (i32 (MVE_VMAXVs16 (t2MOVi32imm (i32 -32768)), $src))>;
def : Pat<(i32 (vecreduce_smin (v4i32 MQPR:$src))),
             (i32 (MVE_VMINVs32 (t2MVNi (i32 -2147483648)), $src))>;
   def : Pat<(i32 (vecreduce_umin (v16i8 MQPR:$src))),
             (i32 (MVE_VMINVu8 (t2MOVi (i32 255)), $src))>;
   def : Pat<(i32 (vecreduce_umin (v8i16 MQPR:$src))),
             (i32 (MVE_VMINVu16 (t2MOVi16 (i32 65535)), $src))>;
   def : Pat<(i32 (vecreduce_umin (v4i32 MQPR:$src))),
             (i32 (MVE_VMINVu32 (t2MOVi (i32 4294967295)), $src))>;
 
+  def : Pat<(i32 (ARMVMINVu (i32 rGPR:$x), (v16i8 MQPR:$src))),
+            (i32 (MVE_VMINVu8 $x, $src))>;
+  def : Pat<(i32 (ARMVMINVu (i32 rGPR:$x), (v8i16 MQPR:$src))),
+            (i32 (MVE_VMINVu16 $x, $src))>;
+  def : Pat<(i32 (ARMVMINVu (i32 rGPR:$x), (v4i32 MQPR:$src))),
+            (i32 (MVE_VMINVu32 $x, $src))>;
+  def : Pat<(i32 (ARMVMINVs (i32 rGPR:$x), (v16i8 MQPR:$src))),
+            (i32 (MVE_VMINVs8 $x, $src))>;
+  def : Pat<(i32 (ARMVMINVs (i32 rGPR:$x), (v8i16 MQPR:$src))),
+            (i32 (MVE_VMINVs16 $x, $src))>;
+  def : Pat<(i32 (ARMVMINVs (i32 rGPR:$x), (v4i32 MQPR:$src))),
+            (i32 (MVE_VMINVs32 $x, $src))>;
+
+  def : Pat<(i32 (ARMVMAXVu (i32 rGPR:$x), (v16i8 MQPR:$src))),
+            (i32 (MVE_VMAXVu8 $x, $src))>;
+  def : Pat<(i32 (ARMVMAXVu (i32 rGPR:$x), (v8i16 MQPR:$src))),
+            (i32 (MVE_VMAXVu16 $x, $src))>;
+  def : Pat<(i32 (ARMVMAXVu (i32 rGPR:$x), (v4i32 MQPR:$src))),
+            (i32 (MVE_VMAXVu32 $x, $src))>;
+  def : Pat<(i32 (ARMVMAXVs (i32 rGPR:$x), (v16i8 MQPR:$src))),
+            (i32 (MVE_VMAXVs8 $x, $src))>;
+  def : Pat<(i32 (ARMVMAXVs (i32 rGPR:$x), (v8i16 MQPR:$src))),
+            (i32 (MVE_VMAXVs16 $x, $src))>;
+  def : Pat<(i32 (ARMVMAXVs (i32 rGPR:$x), (v4i32 MQPR:$src))),
+            (i32 (MVE_VMAXVs32 $x, $src))>;
+
 }
 
 multiclass MVE_VMINMAXAV_ty<string iname, bit isMin, string intrBaseName> {
   defm s8 : MVE_VMINMAXV_p<iname, 0, isMin, MVE_v16s8, intrBaseName>;
   defm s16: MVE_VMINMAXV_p<iname, 0, isMin, MVE_v8s16, intrBaseName>;
   defm s32: MVE_VMINMAXV_p<iname, 0, isMin, MVE_v4s32, intrBaseName>;
 }
 

llvm/test/CodeGen/Thumb2/mve-vmaxv-vminv-scalar.ll

+; RUN: llc -mtriple=thumbv8.1m.main-none-none-eabi -mattr=+mve.fp %s -o - | FileCheck %s
+
+define arm_aapcs_vfpcc i8 @uminv16i8(<16 x i8> %vec, i8 %min) {
+; CHECK-LABEL: uminv16i8:
+; CHECK:       @ %bb.0:
+; CHECK-NEXT:    vminv.u8 r0, q0
+; CHECK-NEXT:    bx lr
+  %x = call i8 @llvm.experimental.vector.reduce.umin.v16i8(<16 x i8> %vec)
+  %cmp = icmp ult i8 %x, %min
+  %1 = select i1 %cmp, i8 %x, i8 %min
+  ret i8 %1
+}
+
+define arm_aapcs_vfpcc i16 @uminv8i16(<8 x i16> %vec, i16 %min) {
+; CHECK-LABEL: uminv8i16:
+; CHECK:       @ %bb.0:
+; CHECK-NEXT:    vminv.u16 r0, q0
+; CHECK-NEXT:    bx lr
+  %x = call i16 @llvm.experimental.vector.reduce.umin.v8i16(<8 x i16> %vec)
+  %cmp = icmp ult i16 %x, %min
+  %1 = select i1 %cmp, i16 %x, i16 %min
+  ret i16 %1
+}
+
+define arm_aapcs_vfpcc i32 @uminv4i32(<4 x i32> %vec, i32 %min) {
+; CHECK-LABEL: uminv4i32:
+; CHECK:       @ %bb.0:
+; CHECK-NEXT:    vminv.u32 r0, q0
+; CHECK-NEXT:    bx lr
+  %x = call i32 @llvm.experimental.vector.reduce.umin.v4i32(<4 x i32> %vec)
+  %cmp = icmp ult i32 %x, %min
+  %1 = select i1 %cmp, i32 %x, i32 %min
+  ret i32 %1
+}
+
+define arm_aapcs_vfpcc i8 @sminv16i8(<16 x i8> %vec, i8 %min) {
+; CHECK-LABEL: sminv16i8:
+; CHECK:       @ %bb.0:
+; CHECK-NEXT:    vminv.s8 r0, q0
+; CHECK-NEXT:    bx lr
+  %x = call i8 @llvm.experimental.vector.reduce.smin.v16i8(<16 x i8> %vec)
+  %cmp = icmp slt i8 %x, %min
+  %1 = select i1 %cmp, i8 %x, i8 %min
+  ret i8 %1
+}
+
+define arm_aapcs_vfpcc i16 @sminv8i16(<8 x i16> %vec, i16 %min) {
+; CHECK-LABEL: sminv8i16:
+; CHECK:       @ %bb.0:
+; CHECK-NEXT:    vminv.s16 r0, q0
+; CHECK-NEXT:    bx lr
+  %x = call i16 @llvm.experimental.vector.reduce.smin.v8i16(<8 x i16> %vec)
+  %cmp = icmp slt i16 %x, %min
+  %1 = select i1 %cmp, i16 %x, i16 %min
+  ret i16 %1
+}
+
+define arm_aapcs_vfpcc i32 @sminv4i32(<4 x i32> %vec, i32 %min) {
+; CHECK-LABEL: sminv4i32:
+; CHECK:       @ %bb.0:
+; CHECK-NEXT:    vminv.s32 r0, q0
+; CHECK-NEXT:    bx lr
+  %x = call i32 @llvm.experimental.vector.reduce.smin.v4i32(<4 x i32> %vec)
+  %cmp = icmp slt i32 %x, %min
+  %1 = select i1 %cmp, i32 %x, i32 %min
+  ret i32 %1
+}
+
+define arm_aapcs_vfpcc i8 @umaxv16i8(<16 x i8> %vec, i8 %max) {
+; CHECK-LABEL: umaxv16i8:
+; CHECK:       @ %bb.0:
+; CHECK-NEXT:    vmaxv.u8 r0, q0
+; CHECK-NEXT:    bx lr
+  %x = call i8 @llvm.experimental.vector.reduce.umax.v16i8(<16 x i8> %vec)
+  %cmp = icmp ugt i8 %x, %max
+  %1 = select i1 %cmp, i8 %x, i8 %max
+  ret i8 %1
+}
+
+define arm_aapcs_vfpcc i16 @umaxv8i16(<8 x i16> %vec, i16 %max) {
+; CHECK-LABEL: umaxv8i16:
+; CHECK:       @ %bb.0:
+; CHECK-NEXT:    vmaxv.u16 r0, q0
+; CHECK-NEXT:    bx lr
+  %x = call i16 @llvm.experimental.vector.reduce.umax.v8i16(<8 x i16> %vec)
+  %cmp = icmp ugt i16 %x, %max
+  %1 = select i1 %cmp, i16 %x, i16 %max
+  ret i16 %1
+}
+
+define arm_aapcs_vfpcc i32 @umaxv4i32(<4 x i32> %vec, i32 %max) {
+; CHECK-LABEL: umaxv4i32:
+; CHECK:       @ %bb.0:
+; CHECK-NEXT:    vmaxv.u32 r0, q0
+; CHECK-NEXT:    bx lr
+  %x = call i32 @llvm.experimental.vector.reduce.umax.v4i32(<4 x i32> %vec)
+  %cmp = icmp ugt i32 %x, %max
+  %1 = select i1 %cmp, i32 %x, i32 %max
+  ret i32 %1
+}
+
+define arm_aapcs_vfpcc i8 @smaxv16i8(<16 x i8> %vec, i8 %max) {
+; CHECK-LABEL: smaxv16i8:
+; CHECK:       @ %bb.0:
+; CHECK-NEXT:    vmaxv.s8 r0, q0
+; CHECK-NEXT:    bx lr
+  %x = call i8 @llvm.experimental.vector.reduce.smax.v16i8(<16 x i8> %vec)
+  %cmp = icmp sgt i8 %x, %max
+  %1 = select i1 %cmp, i8 %x, i8 %max
+  ret i8 %1
+}
+
+define arm_aapcs_vfpcc i16 @smaxv8i16(<8 x i16> %vec, i16 %max) {
+; CHECK-LABEL: smaxv8i16:
+; CHECK:       @ %bb.0:
+; CHECK-NEXT:    vmaxv.s16 r0, q0
+; CHECK-NEXT:    bx lr
+  %x = call i16 @llvm.experimental.vector.reduce.smax.v8i16(<8 x i16> %vec)
+  %cmp = icmp sgt i16 %x, %max
+  %1 = select i1 %cmp, i16 %x, i16 %max
+  ret i16 %1
+}
+
+define arm_aapcs_vfpcc i32 @smaxv4i32(<4 x i32> %vec, i32 %max) {
+; CHECK-LABEL: smaxv4i32:
+; CHECK:       @ %bb.0:
+; CHECK-NEXT:    vmaxv.s32 r0, q0
+; CHECK-NEXT:    bx lr
+  %x = call i32 @llvm.experimental.vector.reduce.smax.v4i32(<4 x i32> %vec)
+  %cmp = icmp sgt i32 %x, %max
+  %1 = select i1 %cmp, i32 %x, i32 %max
+  ret i32 %1
+}
+
+define arm_aapcs_vfpcc i8 @commute_uminv16i8(<16 x i8> %vec, i8 %min) {
+; CHECK-LABEL: commute_uminv16i8:
+; CHECK:       @ %bb.0:
+; CHECK-NEXT:    vminv.u8 r0, q0
+; CHECK-NEXT:    bx lr
+  %x = call i8 @llvm.experimental.vector.reduce.umin.v16i8(<16 x i8> %vec)
+  %cmp = icmp ult i8 %min, %x
+  %1 = select i1 %cmp, i8 %min, i8 %x
+  ret i8 %1
+}
+
+define arm_aapcs_vfpcc i16 @commute_uminv8i16(<8 x i16> %vec, i16 %min) {
+; CHECK-LABEL: commute_uminv8i16:
+; CHECK:       @ %bb.0:
+; CHECK-NEXT:    vminv.u16 r0, q0
+; CHECK-NEXT:    bx lr
+  %x = call i16 @llvm.experimental.vector.reduce.umin.v8i16(<8 x i16> %vec)
+  %cmp = icmp ult i16 %min, %x
+  %1 = select i1 %cmp, i16 %min, i16 %x
+  ret i16 %1
+}
+
+define arm_aapcs_vfpcc i32 @commute_uminv4i32(<4 x i32> %vec, i32 %min) {
+; CHECK-LABEL: commute_uminv4i32:
+; CHECK:       @ %bb.0:
+; CHECK-NEXT:    vminv.u32 r0, q0
+; CHECK-NEXT:    bx lr
+  %x = call i32 @llvm.experimental.vector.reduce.umin.v4i32(<4 x i32> %vec)
+  %cmp = icmp ult i32 %min, %x
+  %1 = select i1 %cmp, i32 %min, i32 %x
+  ret i32 %1
+}
+
+define arm_aapcs_vfpcc i8 @commute_sminv16i8(<16 x i8> %vec, i8 %min) {
+; CHECK-LABEL: commute_sminv16i8:
+; CHECK:       @ %bb.0:
+; CHECK-NEXT:    vminv.s8 r0, q0
+; CHECK-NEXT:    bx lr
+  %x = call i8 @llvm.experimental.vector.reduce.smin.v16i8(<16 x i8> %vec)
+  %cmp = icmp slt i8 %min, %x
+  %1 = select i1 %cmp, i8 %min, i8 %x
+  ret i8 %1
+}
+
+define arm_aapcs_vfpcc i16 @commute_sminv8i16(<8 x i16> %vec, i16 %min) {
+; CHECK-LABEL: commute_sminv8i16:
+; CHECK:       @ %bb.0:
+; CHECK-NEXT:    vminv.s16 r0, q0
+; CHECK-NEXT:    bx lr
+  %x = call i16 @llvm.experimental.vector.reduce.smin.v8i16(<8 x i16> %vec)
+  %cmp = icmp slt i16 %min, %x
+  %1 = select i1 %cmp, i16 %min, i16 %x
+  ret i16 %1
+}
+
+define arm_aapcs_vfpcc i32 @commute_sminv4i32(<4 x i32> %vec, i32 %min) {
+; CHECK-LABEL: commute_sminv4i32:
+; CHECK:       @ %bb.0:
+; CHECK-NEXT:    vminv.s32 r0, q0
+; CHECK-NEXT:    bx lr
+  %x = call i32 @llvm.experimental.vector.reduce.smin.v4i32(<4 x i32> %vec)
+  %cmp = icmp slt i32 %min, %x
+  %1 = select i1 %cmp, i32 %min, i32 %x
+  ret i32 %1
+}
+
+define arm_aapcs_vfpcc i8 @commute_umaxv16i8(<16 x i8> %vec, i8 %max) {
+; CHECK-LABEL: commute_umaxv16i8:
+; CHECK:       @ %bb.0:
+; CHECK-NEXT:    vmaxv.u8 r0, q0
+; CHECK-NEXT:    bx lr
+  %x = call i8 @llvm.experimental.vector.reduce.umax.v16i8(<16 x i8> %vec)
+  %cmp = icmp ugt i8 %max, %x
+  %1 = select i1 %cmp, i8 %max, i8 %x
+  ret i8 %1
+}
+
+define arm_aapcs_vfpcc i16 @commute_umaxv8i16(<8 x i16> %vec, i16 %max) {
+; CHECK-LABEL: commute_umaxv8i16:
+; CHECK:       @ %bb.0:
+; CHECK-NEXT:    vmaxv.u16 r0, q0
+; CHECK-NEXT:    bx lr
+  %x = call i16 @llvm.experimental.vector.reduce.umax.v8i16(<8 x i16> %vec)
+  %cmp = icmp ugt i16 %max, %x
+  %1 = select i1 %cmp, i16 %max, i16 %x
+  ret i16 %1
+}
+
+define arm_aapcs_vfpcc i32 @commute_umaxv4i32(<4 x i32> %vec, i32 %max) {
+; CHECK-LABEL: commute_umaxv4i32:
+; CHECK:       @ %bb.0:
+; CHECK-NEXT:    vmaxv.u32 r0, q0
+; CHECK-NEXT:    bx lr
+  %x = call i32 @llvm.experimental.vector.reduce.umax.v4i32(<4 x i32> %vec)
+  %cmp = icmp ugt i32 %max, %x
+  %1 = select i1 %cmp, i32 %max, i32 %x
+  ret i32 %1
+}
+
+define arm_aapcs_vfpcc i8 @commute_smaxv16i8(<16 x i8> %vec, i8 %max) {
+; CHECK-LABEL: commute_smaxv16i8:
+; CHECK:       @ %bb.0:
+; CHECK-NEXT:    vmaxv.s8 r0, q0
+; CHECK-NEXT:    bx lr
+  %x = call i8 @llvm.experimental.vector.reduce.smax.v16i8(<16 x i8> %vec)
+  %cmp = icmp sgt i8 %max, %x
+  %1 = select i1 %cmp, i8 %max, i8 %x
+  ret i8 %1
+}
+
+define arm_aapcs_vfpcc i16 @commute_smaxv8i16(<8 x i16> %vec, i16 %max) {
+; CHECK-LABEL: commute_smaxv8i16:
+; CHECK:       @ %bb.0:
+; CHECK-NEXT:    vmaxv.s16 r0, q0
+; CHECK-NEXT:    bx lr
+  %x = call i16 @llvm.experimental.vector.reduce.smax.v8i16(<8 x i16> %vec)
+  %cmp = icmp sgt i16 %max, %x
+  %1 = select i1 %cmp, i16 %max, i16 %x
+  ret i16 %1
+}
+
+define arm_aapcs_vfpcc i32 @commute_smaxv4i32(<4 x i32> %vec, i32 %max) {
+; CHECK-LABEL: commute_smaxv4i32:
+; CHECK:       @ %bb.0:
+; CHECK-NEXT:    vmaxv.s32 r0, q0
+; CHECK-NEXT:    bx lr
+  %x = call i32 @llvm.experimental.vector.reduce.smax.v4i32(<4 x i32> %vec)
+  %cmp = icmp sgt i32 %max, %x
+  %1 = select i1 %cmp, i32 %max, i32 %x
+  ret i32 %1
+}
+
+define arm_aapcs_vfpcc i8 @mismatch_smaxv16i8(<16 x i8> %vec, i8 %max) {
+; CHECK-LABEL: mismatch_smaxv16i8:
+; CHECK:       @ %bb.0:
+; CHECK-NEXT:    mvn r1, #127
+; CHECK-NEXT:    sxtb r3, r0
+; CHECK-NEXT:    vmaxv.s8 r1, q0
+; CHECK-NEXT:    sxtb r2, r1
+; CHECK-NEXT:    cmp r2, r3
+; CHECK-NEXT:    csel r0, r0, r1, gt
+; CHECK-NEXT:    bx lr
+  %x = call i8 @llvm.experimental.vector.reduce.smax.v16i8(<16 x i8> %vec)
+  %cmp = icmp sgt i8 %x, %max
+  %1 = select i1 %cmp, i8 %max, i8 %x
+  ret i8 %1
+}
+
+define arm_aapcs_vfpcc i8 @mismatch2_smaxv16i8(<16 x i8> %vec, i8 %max) {
+; CHECK-LABEL: mismatch2_smaxv16i8:
+; CHECK:       @ %bb.0:
+; CHECK-NEXT:    mvn r1, #127
+; CHECK-NEXT:    sxtb r3, r0
+; CHECK-NEXT:    vmaxv.s8 r1, q0
+; CHECK-NEXT:    sxtb r2, r1
+; CHECK-NEXT:    cmp r3, r2
+; CHECK-NEXT:    csel r0, r1, r0, gt
+; CHECK-NEXT:    bx lr
+  %x = call i8 @llvm.experimental.vector.reduce.smax.v16i8(<16 x i8> %vec)
+  %cmp = icmp sgt i8 %max, %x
+  %1 = select i1 %cmp, i8 %x, i8 %max
+  ret i8 %1
+}
+
+declare i8 @llvm.experimental.vector.reduce.umin.v16i8(<16 x i8>)
+
+declare i16 @llvm.experimental.vector.reduce.umin.v8i16(<8 x i16>)
+
+declare i32 @llvm.experimental.vector.reduce.umin.v4i32(<4 x i32>)
+
+declare i8 @llvm.experimental.vector.reduce.smin.v16i8(<16 x i8>)
+
+declare i16 @llvm.experimental.vector.reduce.smin.v8i16(<8 x i16>)
+
+declare i32 @llvm.experimental.vector.reduce.smin.v4i32(<4 x i32>)
+
+declare i8 @llvm.experimental.vector.reduce.umax.v16i8(<16 x i8>)
+
+declare i16 @llvm.experimental.vector.reduce.umax.v8i16(<8 x i16>)
+
+declare i32 @llvm.experimental.vector.reduce.umax.v4i32(<4 x i32>)
+
+declare i8 @llvm.experimental.vector.reduce.smax.v16i8(<16 x i8>)
+
+declare i16 @llvm.experimental.vector.reduce.smax.v8i16(<8 x i16>)
+
+declare i32 @llvm.experimental.vector.reduce.smax.v4i32(<4 x i32>)