diff --git a/Lib/test/test_dis.py b/Lib/test/test_dis.py
index 8597b8f14ac058..8c29523874bde8 100644
--- a/Lib/test/test_dis.py
+++ b/Lib/test/test_dis.py
@@ -1029,28 +1029,31 @@ def expected(count, w):
             s = ['''\
   1        %*d RESUME                   0
 
+''' % (w, 0)]
+            s += ['''\
   2        %*d LOAD_FAST                0 (x)
            %*d LOAD_CONST               1 (1)
            %*d BINARY_OP                0 (+)
-''' % (w, 0, w, 2, w, 4, w, 6)]
+''' % (w, 2, w, 4, w, 6)]
             s += ['''\
-           %*d STORE_FAST_LOAD_FAST     0 (x, x)
+           %*d POP_TOP
+           %*d LOAD_FAST                0 (x)
            %*d LOAD_CONST               1 (1)
            %*d BINARY_OP                0 (+)
-''' % (w, 8*i + 10, w, 8*i + 12, w, 8*i + 14)
-                 for i in range(count-1)]
+''' % (w, 10*i + 10, w, 10*i + 12, w, 10*i + 14, w, 10*i + 16)
+                  for i in range(count-1)]
             s += ['''\
            %*d STORE_FAST               0 (x)
 
   3        %*d LOAD_FAST                0 (x)
            %*d RETURN_VALUE
-''' % (w, 8*count + 2, w, 8*count + 4, w, 8*count + 6)]
+''' % (w, 10*count, w, 10*count + 2, w, 10*count + 4)]
             return ''.join(s)
 
         for i in range(1, 5):
             self.do_disassembly_test(func(i), expected(i, 4), True)
-        self.do_disassembly_test(func(1200), expected(1200, 4), True)
-        self.do_disassembly_test(func(1300), expected(1300, 5), True)
+        self.do_disassembly_test(func(999), expected(999, 4), True)
+        self.do_disassembly_test(func(1000), expected(1000, 5), True)
 
     def test_disassemble_str(self):
         self.do_disassembly_test(expr_str, dis_expr_str)
diff --git a/Lib/test/test_peepholer.py b/Lib/test/test_peepholer.py
index 82b0b50d0ea437..3a6af0250e690c 100644
--- a/Lib/test/test_peepholer.py
+++ b/Lib/test/test_peepholer.py
@@ -1104,6 +1104,25 @@ def test_dead_store_elimination_in_same_lineno(self):
         ]
         self.cfg_optimization_test(insts, expected_insts, consts=list(range(3)), nlocals=1)
 
+    def test_dead_store_elimination_in_same_lineno_wider(self):
+        insts = [
+            ('LOAD_CONST', 0, 1),
+            ('LOAD_CONST', 1, 2),
+            ('LOAD_CONST', 2, 3),
+            ('STORE_FAST', 1, 4),
+            ('STORE_FAST', 0, 4),
+            ('STORE_FAST', 1, 4),
+            ('RETURN_VALUE', 5)
+        ]
+        expected_insts = [
+            ('LOAD_CONST', 0, 1),
+            ('LOAD_CONST', 1, 2),
+            ('NOP', 0, 3),
+            ('STORE_FAST_STORE_FAST', 1, 4),
+            ('RETURN_VALUE', 5)
+        ]
+        self.cfg_optimization_test(insts, expected_insts, consts=list(range(3)), nlocals=1)
+
     def test_no_dead_store_elimination_in_different_lineno(self):
         insts = [
             ('LOAD_CONST', 0, 1),
diff --git a/Python/flowgraph.c b/Python/flowgraph.c
index e159a4356dfe46..d2c32988f3ca29 100644
--- a/Python/flowgraph.c
+++ b/Python/flowgraph.c
@@ -1303,6 +1303,30 @@ apply_static_swaps(basicblock *block, int i)
     }
 }
 
+static void
+remove_redundant_deadstore(basicblock *bb)
+{
+    for (int i = bb->b_iused - 1; i >= 0; i--) {
+        cfg_instr *inst = &bb->b_instr[i];
+        // Optimize only if the opcode is STORE_FAST and the lineno is same.
+        // The number of STORE_FAST will be decreased if the optimization is success.
+        // So most of case could be skipped.
+        if (inst->i_opcode == STORE_FAST) {
+            for (int j = i - 1; j >= 0; j--) {
+                cfg_instr *prev = &bb->b_instr[j];
+                if (prev->i_loc.lineno != inst->i_loc.lineno) {
+                    // Invariant condition: lineno is always monotonically increasing.
+                    break;
+                }
+                if (prev->i_opcode == STORE_FAST && prev->i_oparg == inst->i_oparg) {
+                    prev->i_opcode = POP_TOP;
+                    prev->i_oparg = 0;
+                }
+            }
+        }
+    }
+}
+
 static int
 optimize_basic_block(PyObject *const_cache, basicblock *bb, PyObject *consts)
 {
@@ -1314,6 +1338,9 @@ optimize_basic_block(PyObject *const_cache, basicblock *bb, PyObject *consts)
     int opcode = 0;
     int oparg = 0;
     int nextop = 0;
+
+    remove_redundant_deadstore(bb);
+
     for (int i = 0; i < bb->b_iused; i++) {
         cfg_instr *inst = &bb->b_instr[i];
         bool is_copy_of_load_const = (opcode == LOAD_CONST &&
@@ -1491,14 +1518,6 @@ optimize_basic_block(PyObject *const_cache, basicblock *bb, PyObject *consts)
                     */
                 }
                 break;
-            case STORE_FAST:
-                if (opcode == nextop &&
-                    oparg == bb->b_instr[i+1].i_oparg &&
-                    bb->b_instr[i].i_loc.lineno == bb->b_instr[i+1].i_loc.lineno) {
-                    bb->b_instr[i].i_opcode = POP_TOP;
-                    bb->b_instr[i].i_oparg = 0;
-                }
-                break;
             case SWAP:
                 if (oparg == 1) {
                     INSTR_SET_OP0(inst, NOP);