diff --git a/src/cmd/compile/internal/ssacompile/memcombine.go b/src/cmd/compile/internal/ssacompile/memcombine.go
index 33898ff0098ddf..0ebd8f539ba85f 100644
--- a/src/cmd/compile/internal/ssacompile/memcombine.go
+++ b/src/cmd/compile/internal/ssacompile/memcombine.go
@@ -51,8 +51,8 @@ func memcombineLoads(f *ssa.Func, ptrAlignments []int8) {
 			if v.Op != ssaop.OpOr16 && v.Op != ssaop.OpOr32 && v.Op != ssaop.OpOr64 {
 				continue
 			}
-			if mark.Contains(v.ID) {
-				// marked - means it is not the root of an OR tree
+			if v.Uses == 1 && mark.Contains(v.ID) {
+				// A single-use argument belongs to its parent's OR tree.
 				continue
 			}
 			// Add the OR tree rooted at v to the order.
@@ -63,7 +63,8 @@ func memcombineLoads(f *ssa.Func, ptrAlignments []int8) {
 				x := order[i]
 				for j := 0; j < 2; j++ {
 					a := x.Args[j]
-					if a.Op == ssaop.OpOr16 || a.Op == ssaop.OpOr32 || a.Op == ssaop.OpOr64 {
+					// combineLoads cannot combine through shared values.
+					if a.Uses == 1 && (a.Op == ssaop.OpOr16 || a.Op == ssaop.OpOr32 || a.Op == ssaop.OpOr64) {
 						order = append(order, a)
 					}
 				}
diff --git a/src/cmd/compile/internal/ssacompile/memcombine_test.go b/src/cmd/compile/internal/ssacompile/memcombine_test.go
new file mode 100644
index 00000000000000..20b9d639fe3e1f
--- /dev/null
+++ b/src/cmd/compile/internal/ssacompile/memcombine_test.go
@@ -0,0 +1,41 @@
+// Copyright 2026 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package ssacompile
+
+import (
+	"fmt"
+	"testing"
+
+	"cmd/compile/internal/ssa"
+	"cmd/compile/internal/ssa/block"
+	"cmd/compile/internal/ssa/ssaop"
+	"cmd/internal/src"
+)
+
+func BenchmarkMemcombineSharedOR(b *testing.B) {
+	for _, n := range []int{8, 16, 22} {
+		b.Run(fmt.Sprint(n), func(b *testing.B) {
+			c := testConfig(b)
+			f := c.config.NewFunc(c.Frontend(), new(ssa.Cache))
+			entry := f.NewBlock(block.BlockExit)
+			f.Entry = entry
+			typ := c.config.Types.UInt64
+			x := entry.NewValue0(src.NoXPos, ssaop.OpArg, typ)
+			y := entry.NewValue0(src.NoXPos, ssaop.OpArg, typ)
+			v := entry.NewValue2(src.NoXPos, ssaop.OpOr64, typ, x, y)
+			// Each level uses the previous result through two parents.
+			// Expanding this graph as a tree takes exponential work.
+			for range n {
+				a := entry.NewValue2(src.NoXPos, ssaop.OpOr64, typ, v, x)
+				b := entry.NewValue2(src.NoXPos, ssaop.OpOr64, typ, v, y)
+				v = entry.NewValue2(src.NoXPos, ssaop.OpOr64, typ, a, b)
+			}
+			b.ReportAllocs()
+			for b.Loop() {
+				memcombineLoads(f, nil)
+			}
+		})
+	}
+}
diff --git a/test/codegen/memcombine.go b/test/codegen/memcombine.go
index 9e863aab5ce0d2..01404081cfcb9d 100644
--- a/test/codegen/memcombine.go
+++ b/test/codegen/memcombine.go
@@ -68,6 +68,14 @@ func load_le16(b []byte) uint16 {
 	return binary.LittleEndian.Uint16(b)
 }
 
+// A combined load that is shared by two OR expressions must still be found.
+func load_le16_shared(b []byte, x, y uint16) (uint16, uint16) {
+	// amd64:`MOVWLZX \(.*\),` -`MOVB`
+	// arm64:`MOVHU \(R[0-9]+\),` -`MOVB`
+	v := uint16(b[0]) | uint16(b[1])<<8
+	return v | x, v | y
+}
+
 func load_le16_idx(b []byte, idx int) uint16 {
 	// amd64:`MOVWLZX \(.*\),` -`MOVB` -`OR`
 	// ppc64le:`MOVHZ ` -`MOVBZ`