[ARM] Match MVE vqdmulh

davemgreen · davemgreen · commit d14db8c8dc98 · 2020-10-30T13:34:27.000Z
This adds ISel matching for a form of VQDMULH. There are several ir patterns that we could match to that instruction, this one is for: min(ashr(mul(sext(a), sext(b)), 7), 127) Which is what llvm will optimize to once it has removed the max that usually makes up the min/max saturate pattern, as in this case the compare will always be false. The additional complication to match i32 patterns (which extend into an i64) is that the min will be a vselect/setcc, as vmin is not supported for i64 vectors. Tablegen patterns have also been updated to attempt to reuse the MVE_TwoOpPattern patterns. Differential Revision: https://reviews.llvm.org/D90096
diff --git a/llvm/lib/Target/ARM/ARMISelLowering.cpp b/llvm/lib/Target/ARM/ARMISelLowering.cpp
@@ -1718,6 +1718,7 @@ const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
   case ARMISD::VCVTL:         return "ARMISD::VCVTL";
   case ARMISD::VMULLs:        return "ARMISD::VMULLs";
   case ARMISD::VMULLu:        return "ARMISD::VMULLu";
+  case ARMISD::VQDMULH:       return "ARMISD::VQDMULH";
   case ARMISD::VADDVs:        return "ARMISD::VADDVs";
   case ARMISD::VADDVu:        return "ARMISD::VADDVu";
   case ARMISD::VADDVps:       return "ARMISD::VADDVps";
@@ -12206,9 +12207,93 @@ static SDValue PerformSELECTCombine(SDNode *N,
   return Reduction;
 }
 
+// A special combine for the vqdmulh family of instructions. This is one of the
+// potential set of patterns that could patch this instruction. The base pattern
+// you would expect to be min(max(ashr(mul(mul(sext(x), 2), sext(y)), 16))).
+// This matches the different min(max(ashr(mul(mul(sext(x), sext(y)), 2), 16))),
+// which llvm will have optimized to min(ashr(mul(sext(x), sext(y)), 15))) as
+// the max is unnecessary.
+static SDValue PerformVQDMULHCombine(SDNode *N, SelectionDAG &DAG) {
+  EVT VT = N->getValueType(0);
+  SDValue Shft;
+  ConstantSDNode *Clamp;
+
+  if (N->getOpcode() == ISD::SMIN) {
+    Shft = N->getOperand(0);
+    Clamp = isConstOrConstSplat(N->getOperand(1));
+  } else if (N->getOpcode() == ISD::VSELECT) {
+    // Detect a SMIN, which for an i64 node will be a vselect/setcc, not a smin.
+    SDValue Cmp = N->getOperand(0);
+    if (Cmp.getOpcode() != ISD::SETCC ||
+        cast<CondCodeSDNode>(Cmp.getOperand(2))->get() != ISD::SETLT ||
+        Cmp.getOperand(0) != N->getOperand(1) ||
+        Cmp.getOperand(1) != N->getOperand(2))
+      return SDValue();
+    Shft = N->getOperand(1);
+    Clamp = isConstOrConstSplat(N->getOperand(2));
+  } else
+    return SDValue();
+
+  if (!Clamp)
+    return SDValue();
+
+  MVT ScalarType;
+  int ShftAmt = 0;
+  switch (Clamp->getSExtValue()) {
+  case (1 << 7) - 1:
+    ScalarType = MVT::i8;
+    ShftAmt = 7;
+    break;
+  case (1 << 15) - 1:
+    ScalarType = MVT::i16;
+    ShftAmt = 15;
+    break;
+  case (1ULL << 31) - 1:
+    ScalarType = MVT::i32;
+    ShftAmt = 31;
+    break;
+  default:
+    return SDValue();
+  }
+
+  if (Shft.getOpcode() != ISD::SRA)
+    return SDValue();
+  ConstantSDNode *N1 = isConstOrConstSplat(Shft.getOperand(1));
+  if (!N1 || N1->getSExtValue() != ShftAmt)
+    return SDValue();
+
+  SDValue Mul = Shft.getOperand(0);
+  if (Mul.getOpcode() != ISD::MUL)
+    return SDValue();
+
+  SDValue Ext0 = Mul.getOperand(0);
+  SDValue Ext1 = Mul.getOperand(1);
+  if (Ext0.getOpcode() != ISD::SIGN_EXTEND ||
+      Ext1.getOpcode() != ISD::SIGN_EXTEND)
+    return SDValue();
+  EVT VecVT = Ext0.getOperand(0).getValueType();
+  if (VecVT != MVT::v4i32 && VecVT != MVT::v8i16 && VecVT != MVT::v16i8)
+    return SDValue();
+  if (Ext1.getOperand(0).getValueType() != VecVT ||
+      VecVT.getScalarType() != ScalarType ||
+      VT.getScalarSizeInBits() < ScalarType.getScalarSizeInBits() * 2)
+    return SDValue();
+
+  SDLoc DL(Mul);
+  SDValue VQDMULH = DAG.getNode(ARMISD::VQDMULH, DL, VecVT, Ext0.getOperand(0),
+                                Ext1.getOperand(0));
+  return DAG.getNode(ISD::SIGN_EXTEND, DL, VT, VQDMULH);
+}
+
 static SDValue PerformVSELECTCombine(SDNode *N,
                                      TargetLowering::DAGCombinerInfo &DCI,
                                      const ARMSubtarget *Subtarget) {
+  if (!Subtarget->hasMVEIntegerOps())
+    return SDValue();
+
+  if (SDValue V = PerformVQDMULHCombine(N, DCI.DAG))
+    return V;
+
   // Transforms vselect(not(cond), lhs, rhs) into vselect(cond, rhs, lhs).
   //
   // We need to re-implement this optimization here as the implementation in the
@@ -12218,9 +12303,6 @@ static SDValue PerformVSELECTCombine(SDNode *N,
   //
   // 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);
@@ -14582,6 +14664,14 @@ static SDValue PerformSplittingToNarrowingStores(StoreSDNode *St,
     return true;
   };
 
+  // It may be preferable to keep the store unsplit as the trunc may end up
+  // being removed. Check that here.
+  if (Trunc.getOperand(0).getOpcode() == ISD::SMIN) {
+    if (SDValue U = PerformVQDMULHCombine(Trunc.getOperand(0).getNode(), DAG)) {
+      DAG.ReplaceAllUsesWith(Trunc.getOperand(0), U);
+      return SDValue();
+    }
+  }
   if (auto *Shuffle = dyn_cast<ShuffleVectorSDNode>(Trunc->getOperand(0)))
     if (isVMOVNOriginalMask(Shuffle->getMask(), false) ||
         isVMOVNOriginalMask(Shuffle->getMask(), true))
@@ -15555,6 +15645,9 @@ static SDValue PerformMinMaxCombine(SDNode *N, SelectionDAG &DAG,
   if (!ST->hasMVEIntegerOps())
     return SDValue();
 
+  if (SDValue V = PerformVQDMULHCombine(N, DAG))
+    return V;
+
   if (VT != MVT::v4i32 && VT != MVT::v8i16)
     return SDValue();
 
diff --git a/llvm/lib/Target/ARM/ARMISelLowering.h b/llvm/lib/Target/ARM/ARMISelLowering.h
@@ -216,6 +216,8 @@ class VectorType;
       VMULLs,       // ...signed
       VMULLu,       // ...unsigned
 
+      VQDMULH,      // MVE vqdmulh instruction
+
       // MVE reductions
       VADDVs,       // sign- or zero-extend the elements of a vector to i32,
       VADDVu,       //   add them all together, and return an i32 of their sum
diff --git a/llvm/lib/Target/ARM/ARMInstrMVE.td b/llvm/lib/Target/ARM/ARMInstrMVE.td
@@ -1955,28 +1955,26 @@ class MVE_VQxDMULH_Base<string iname, string suffix, bits<2> size, bit rounding,
   let validForTailPredication = 1;
 }
 
+def MVEvqdmulh : SDNode<"ARMISD::VQDMULH", SDTIntBinOp>;
+
 multiclass MVE_VQxDMULH_m<string iname, MVEVectorVTInfo VTI,
-                      SDNode unpred_op, Intrinsic pred_int,
+                      SDNode Op, Intrinsic unpred_int, Intrinsic pred_int,
                       bit rounding> {
   def "" : MVE_VQxDMULH_Base<iname, VTI.Suffix, VTI.Size, rounding>;
   defvar Inst = !cast<Instruction>(NAME);
+  defm : MVE_TwoOpPattern<VTI, Op, pred_int, (? ), Inst>;
 
   let Predicates = [HasMVEInt] in {
-    // Unpredicated multiply
-    def : Pat<(VTI.Vec (unpred_op (VTI.Vec MQPR:$Qm), (VTI.Vec MQPR:$Qn))),
+    // Extra unpredicated multiply intrinsic patterns
+    def : Pat<(VTI.Vec (unpred_int (VTI.Vec MQPR:$Qm), (VTI.Vec MQPR:$Qn))),
               (VTI.Vec (Inst (VTI.Vec MQPR:$Qm), (VTI.Vec MQPR:$Qn)))>;
-
-    // Predicated multiply
-    def : Pat<(VTI.Vec (pred_int (VTI.Vec MQPR:$Qm), (VTI.Vec MQPR:$Qn),
-                            (VTI.Pred VCCR:$mask), (VTI.Vec MQPR:$inactive))),
-              (VTI.Vec (Inst (VTI.Vec MQPR:$Qm), (VTI.Vec MQPR:$Qn),
-                            ARMVCCThen, (VTI.Pred VCCR:$mask),
-                            (VTI.Vec MQPR:$inactive)))>;
   }
 }
 
 multiclass MVE_VQxDMULH<string iname, MVEVectorVTInfo VTI, bit rounding>
-  : MVE_VQxDMULH_m<iname, VTI, !if(rounding, int_arm_mve_vqrdmulh,
+  : MVE_VQxDMULH_m<iname, VTI, !if(rounding, null_frag,
+                                             MVEvqdmulh),
+                               !if(rounding, int_arm_mve_vqrdmulh,
                                              int_arm_mve_vqdmulh),
                                !if(rounding, int_arm_mve_qrdmulh_predicated,
                                              int_arm_mve_qdmulh_predicated),
@@ -5492,18 +5490,18 @@ class MVE_VxxMUL_qr<string iname, string suffix,
 }
 
 multiclass MVE_VxxMUL_qr_m<string iname, MVEVectorVTInfo VTI, bit bit_28,
-                           Intrinsic int_unpred, Intrinsic int_pred> {
+                           PatFrag Op, Intrinsic int_unpred, Intrinsic int_pred> {
   def "" : MVE_VxxMUL_qr<iname, VTI.Suffix, bit_28, VTI.Size>;
-  defm : MVE_vec_scalar_int_pat_m<!cast<Instruction>(NAME), VTI,
-                                  int_unpred, int_pred>;
+  defm : MVE_TwoOpPatternDup<VTI, Op, int_pred, (? ), !cast<Instruction>(NAME)>;
+  defm : MVE_vec_scalar_int_pat_m<!cast<Instruction>(NAME), VTI, int_unpred, int_pred>;
 }
 
 multiclass MVE_VQDMULH_qr_m<MVEVectorVTInfo VTI> :
-  MVE_VxxMUL_qr_m<"vqdmulh", VTI, 0b0,
+  MVE_VxxMUL_qr_m<"vqdmulh", VTI, 0b0, MVEvqdmulh,
                   int_arm_mve_vqdmulh, int_arm_mve_qdmulh_predicated>;
 
 multiclass MVE_VQRDMULH_qr_m<MVEVectorVTInfo VTI> :
-  MVE_VxxMUL_qr_m<"vqrdmulh", VTI, 0b1,
+  MVE_VxxMUL_qr_m<"vqrdmulh", VTI, 0b1, null_frag,
                   int_arm_mve_vqrdmulh, int_arm_mve_qrdmulh_predicated>;
 
 defm MVE_VQDMULH_qr_s8    : MVE_VQDMULH_qr_m<MVE_v16s8>;
diff --git a/llvm/test/CodeGen/Thumb2/mve-vqdmulh.ll b/llvm/test/CodeGen/Thumb2/mve-vqdmulh.ll
