[SDAG] Use shifts if ISD::MUL is illegal when lowering ISD::CTPOP/ISD::VP_CTPOP

We can avoid libcalls.

Fixes #86205

Author: wangpc-pp

llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp

@@ -8710,11 +8710,19 @@ SDValue TargetLowering::expandCTPOP(SDNode *Node, SelectionDAG &DAG) const {
   }
 
   // v = (v * 0x01010101...) >> (Len - 8)
-  SDValue Mask01 =
-      DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x01)), dl, VT);
-  return DAG.getNode(ISD::SRL, dl, VT,
-                     DAG.getNode(ISD::MUL, dl, VT, Op, Mask01),
-                     DAG.getConstant(Len - 8, dl, ShVT));
+  SDValue V;
+  if (isOperationLegalOrCustomOrPromote(ISD::MUL, VT)) {
+    SDValue Mask01 =
+        DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x01)), dl, VT);
+    V = DAG.getNode(ISD::MUL, dl, VT, Op, Mask01);
+  } else {
+    V = Op;
+    for (unsigned Shift = 8; Shift < Len; Shift *= 2)
+      V = DAG.getNode(ISD::ADD, dl, VT, V,
+                      DAG.getNode(ISD::SHL, dl, VT, V,
+                                  DAG.getShiftAmountConstant(Shift, VT, dl)));
+  }
+  return DAG.getNode(ISD::SRL, dl, VT, V, DAG.getConstant(Len - 8, dl, ShVT));
 }
 
 SDValue TargetLowering::expandVPCTPOP(SDNode *Node, SelectionDAG &DAG) const {
@@ -8767,10 +8775,21 @@ SDValue TargetLowering::expandVPCTPOP(SDNode *Node, SelectionDAG &DAG) const {
     return Op;
 
   // v = (v * 0x01010101...) >> (Len - 8)
-  SDValue Mask01 =
-      DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x01)), dl, VT);
-  return DAG.getNode(ISD::VP_LSHR, dl, VT,
-                     DAG.getNode(ISD::VP_MUL, dl, VT, Op, Mask01, Mask, VL),
+  SDValue V;
+  if (isOperationLegalOrCustomOrPromote(ISD::VP_MUL, VT)) {
+    SDValue Mask01 =
+        DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x01)), dl, VT);
+    V = DAG.getNode(ISD::VP_MUL, dl, VT, Op, Mask01, Mask, VL);
+  } else {
+    V = Op;
+    for (unsigned Shift = 8; Shift < Len; Shift *= 2)
+      V = DAG.getNode(ISD::VP_ADD, dl, VT, V,
+                      DAG.getNode(ISD::VP_SHL, dl, VT, V,
+                                  DAG.getShiftAmountConstant(Shift, VT, dl),
+                                  Mask, VL),
+                      Mask, VL);
+  }
+  return DAG.getNode(ISD::VP_LSHR, dl, VT, V,
                      DAG.getConstant(Len - 8, dl, ShVT), Mask, VL);
 }
 

llvm/test/CodeGen/LoongArch/ctlz-cttz-ctpop.ll

@@ -285,10 +285,11 @@ define i32 @test_ctpop_i32(i32 %a) nounwind {
 ; LA64-NEXT:    lu12i.w $a1, 61680
 ; LA64-NEXT:    ori $a1, $a1, 3855
 ; LA64-NEXT:    and $a0, $a0, $a1
-; LA64-NEXT:    lu12i.w $a1, 4112
-; LA64-NEXT:    ori $a1, $a1, 257
-; LA64-NEXT:    mul.d $a0, $a0, $a1
-; LA64-NEXT:    bstrpick.d $a0, $a0, 31, 24
+; LA64-NEXT:    slli.d $a1, $a0, 8
+; LA64-NEXT:    add.d $a0, $a0, $a1
+; LA64-NEXT:    slli.d $a1, $a0, 16
+; LA64-NEXT:    add.d $a0, $a0, $a1
+; LA64-NEXT:    bstrpick.d $a0, $a0, 29, 24
 ; LA64-NEXT:    ret
   %1 = call i32 @llvm.ctpop.i32(i32 %a)
   ret i32 %1