[RISCV] Update combineBinOp_VLToVWBinOp_VL to check the number users after we know...

If we're forming vwadd(u).wv we don't care if the already wide
source operand has additional users. It's not affected by the
transform.

Fixes #59345

diff --git a/llvm/lib/Target/RISCV/RISCVISelLowering.cpp b/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
index f946225..74dc737 100644 (file)
--- a/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
+++ b/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
@@ -9416,8 +9416,6 @@ combineBinOp_VLToVWBinOp_VL(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
         }
       }
     };
-    AppendUsersIfNeeded(LHS);
-    AppendUsersIfNeeded(RHS);
 
     // Control the compile time by limiting the number of node we look at in
     // total.
@@ -9439,6 +9437,13 @@ combineBinOp_VLToVWBinOp_VL(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
         if (Res) {
           Matched = true;
           CombinesToApply.push_back(*Res);
+          // All the inputs that are extended need to be folded, otherwise
+          // we would be leaving the old input (since it is may still be used),
+          // and the new one.
+          if (Res->SExtLHS.has_value())
+            AppendUsersIfNeeded(LHS);
+          if (Res->SExtRHS.has_value())
+            AppendUsersIfNeeded(RHS);
           break;
         }
       }

diff --git a/llvm/test/CodeGen/RISCV/rvv/vwadd-vp.ll b/llvm/test/CodeGen/RISCV/rvv/vwadd-vp.ll
index e2b77b0..3c5f739 100644 (file)
--- a/llvm/test/CodeGen/RISCV/rvv/vwadd-vp.ll
+++ b/llvm/test/CodeGen/RISCV/rvv/vwadd-vp.ll
@@ -6,10 +6,11 @@ define <vscale x 2 x i32> @vwadd_tu(<vscale x 2 x i8> %arg, <vscale x 2 x i32> %
 ; CHECK:       # %bb.0: # %bb
 ; CHECK-NEXT:    slli a0, a0, 32
 ; CHECK-NEXT:    srli a0, a0, 32
-; CHECK-NEXT:    vsetvli zero, a0, e32, m1, ta, ma
-; CHECK-NEXT:    vsext.vf4 v10, v8
+; CHECK-NEXT:    vsetvli zero, a0, e16, mf2, ta, ma
+; CHECK-NEXT:    vsext.vf2 v10, v8
+; CHECK-NEXT:    vwadd.wv v8, v9, v10
 ; CHECK-NEXT:    vsetvli zero, zero, e32, m1, tu, ma
-; CHECK-NEXT:    vadd.vv v9, v9, v10
+; CHECK-NEXT:    vadd.vi v9, v8, 0
 ; CHECK-NEXT:    vmv1r.v v8, v9
 ; CHECK-NEXT:    ret
 bb:
@@ -24,10 +25,11 @@ define <vscale x 2 x i32> @vwaddu_tu(<vscale x 2 x i8> %arg, <vscale x 2 x i32>
 ; CHECK:       # %bb.0: # %bb
 ; CHECK-NEXT:    slli a0, a0, 32
 ; CHECK-NEXT:    srli a0, a0, 32
-; CHECK-NEXT:    vsetvli zero, a0, e32, m1, ta, ma
-; CHECK-NEXT:    vzext.vf4 v10, v8
+; CHECK-NEXT:    vsetvli zero, a0, e16, mf2, ta, ma
+; CHECK-NEXT:    vzext.vf2 v10, v8
+; CHECK-NEXT:    vwaddu.wv v8, v9, v10
 ; CHECK-NEXT:    vsetvli zero, zero, e32, m1, tu, ma
-; CHECK-NEXT:    vadd.vv v9, v9, v10
+; CHECK-NEXT:    vadd.vi v9, v8, 0
 ; CHECK-NEXT:    vmv1r.v v8, v9
 ; CHECK-NEXT:    ret
 bb: