diff --git a/evm/src/all_stark.rs b/evm/src/all_stark.rs
index 038c1b93..fbcd2115 100644
--- a/evm/src/all_stark.rs
+++ b/evm/src/all_stark.rs
@@ -260,6 +260,7 @@ mod tests {
             let mut row: cpu::columns::CpuColumnsView<F> =
                 [F::ZERO; CpuStark::<F, D>::COLUMNS].into();
             row.is_cpu_cycle = F::ONE;
+            row.program_counter = F::from_canonical_usize(i);
             row.opcode = [
                 (logic::columns::IS_AND, 0x16),
                 (logic::columns::IS_OR, 0x17),
@@ -319,10 +320,11 @@ mod tests {
         }
 
         // Pad to a power of two.
-        for _ in cpu_trace_rows.len()..cpu_trace_rows.len().next_power_of_two() {
+        for i in 0..cpu_trace_rows.len().next_power_of_two() - cpu_trace_rows.len() {
             let mut row: cpu::columns::CpuColumnsView<F> =
                 [F::ZERO; CpuStark::<F, D>::COLUMNS].into();
             row.is_cpu_cycle = F::ONE;
+            row.program_counter = F::from_canonical_usize(i + num_logic_rows);
             cpu_stark.generate(row.borrow_mut());
             cpu_trace_rows.push(row.into());
         }
@@ -335,10 +337,6 @@ mod tests {
             let last_row: &mut cpu::columns::CpuColumnsView<F> =
                 cpu_trace_rows[num_rows - 1].borrow_mut();
             last_row.program_counter = halt_label;
-
-            let second_last_row: &mut cpu::columns::CpuColumnsView<F> =
-                cpu_trace_rows[num_rows - 2].borrow_mut();
-            second_last_row.next_program_counter = halt_label;
         }
 
         trace_rows_to_poly_values(cpu_trace_rows)
diff --git a/evm/src/cpu/columns/mod.rs b/evm/src/cpu/columns/mod.rs
index fbf20af2..1f0eff2d 100644
--- a/evm/src/cpu/columns/mod.rs
+++ b/evm/src/cpu/columns/mod.rs
@@ -146,9 +146,6 @@ pub struct CpuColumnsView<T: Copy> {
     /// If CPU cycle: the opcode, broken up into bits in **big-endian** order.
     pub opcode_bits: [T; 8],
 
-    /// If CPU cycle: The program counter for the next instruction.
-    pub next_program_counter: T,
-
     /// Filter. 1 iff a Keccak permutation is computed on this row.
     pub is_keccak: T,
 
diff --git a/evm/src/cpu/control_flow.rs b/evm/src/cpu/control_flow.rs
index cf24afca..90a76d46 100644
--- a/evm/src/cpu/control_flow.rs
+++ b/evm/src/cpu/control_flow.rs
@@ -5,9 +5,38 @@ use plonky2::hash::hash_types::RichField;
 use plonky2::iop::ext_target::ExtensionTarget;
 
 use crate::constraint_consumer::{ConstraintConsumer, RecursiveConstraintConsumer};
-use crate::cpu::columns::CpuColumnsView;
+use crate::cpu::columns::{CpuColumnsView, COL_MAP};
 use crate::cpu::kernel::aggregator::KERNEL;
 
+// TODO: This list is incomplete.
+const NATIVE_INSTRUCTIONS: [usize; 25] = [
+    COL_MAP.is_add,
+    COL_MAP.is_mul,
+    COL_MAP.is_sub,
+    COL_MAP.is_div,
+    COL_MAP.is_sdiv,
+    COL_MAP.is_mod,
+    COL_MAP.is_smod,
+    COL_MAP.is_addmod,
+    COL_MAP.is_mulmod,
+    COL_MAP.is_signextend,
+    COL_MAP.is_lt,
+    COL_MAP.is_gt,
+    COL_MAP.is_slt,
+    COL_MAP.is_sgt,
+    COL_MAP.is_eq,
+    COL_MAP.is_iszero,
+    COL_MAP.is_and,
+    COL_MAP.is_or,
+    COL_MAP.is_xor,
+    COL_MAP.is_not,
+    COL_MAP.is_byte,
+    COL_MAP.is_shl,
+    COL_MAP.is_shr,
+    COL_MAP.is_sar,
+    COL_MAP.is_pop,
+];
+
 fn get_halt_pcs<F: Field>() -> (F, F) {
     let halt_pc0 = KERNEL.global_labels["halt_pc0"];
     let halt_pc1 = KERNEL.global_labels["halt_pc1"];
@@ -26,10 +55,15 @@ pub fn eval_packed_generic<P: PackedField>(
     // Once we start executing instructions, then we continue until the end of the table.
     yield_constr.constraint_transition(lv.is_cpu_cycle * (nv.is_cpu_cycle - P::ONES));
 
-    // If a row is a CPU cycle, then its `next_program_counter` becomes the `program_counter` of the
-    // next row.
-    yield_constr
-        .constraint_transition(lv.is_cpu_cycle * (nv.program_counter - lv.next_program_counter));
+    // If a row is a CPU cycle and executing a native instruction (implemented as a table row; not
+    // microcoded) then the program counter is incremented by 1 to obtain the next row's program
+    // counter.
+    let is_native_instruction: P = NATIVE_INSTRUCTIONS.iter().map(|&col_i| lv[col_i]).sum();
+    yield_constr.constraint_transition(
+        lv.is_cpu_cycle
+            * is_native_instruction
+            * (lv.program_counter - nv.program_counter + P::ONES),
+    );
 
     // If a non-CPU cycle row is followed by a CPU cycle row, then the `program_counter` of the CPU
     // cycle row is 0.
@@ -64,11 +98,15 @@ pub fn eval_ext_circuit<F: RichField + Extendable<D>, const D: usize>(
         yield_constr.constraint_transition(builder, constr);
     }
 
-    // If a row is a CPU cycle, then its `next_program_counter` becomes the `program_counter` of the
-    // next row.
+    // If a row is a CPU cycle and executing a native instruction (implemented as a table row; not
+    // microcoded) then the program counter is incremented by 1 to obtain the next row's program
+    // counter.
     {
-        let constr = builder.sub_extension(nv.program_counter, lv.next_program_counter);
-        let constr = builder.mul_extension(lv.is_cpu_cycle, constr);
+        let is_native_instruction =
+            builder.add_many_extension(NATIVE_INSTRUCTIONS.iter().map(|&col_i| lv[col_i]));
+        let filter = builder.mul_extension(lv.is_cpu_cycle, is_native_instruction);
+        let pc_diff = builder.sub_extension(lv.program_counter, nv.program_counter);
+        let constr = builder.mul_add_extension(filter, pc_diff, filter);
         yield_constr.constraint_transition(builder, constr);
     }