Combine DUP and SWAP (#1254)

* Combine dup and swap flags.

* Add comments

evm/src/cpu/columns/ops.rs

@@ -22,8 +22,7 @@ pub struct OpsColumnsView<T: Copy> {
     pub jumpdest: T,
     pub push0: T,
     pub push: T,
-    pub dup: T,
-    pub swap: T,
+    pub dup_swap: T,
     pub get_context: T,
     pub set_context: T,
     pub mstore_32bytes: T,

evm/src/cpu/control_flow.rs

@@ -8,7 +8,7 @@ use crate::constraint_consumer::{ConstraintConsumer, RecursiveConstraintConsumer
 use crate::cpu::columns::{CpuColumnsView, COL_MAP};
 use crate::cpu::kernel::aggregator::KERNEL;
 
-const NATIVE_INSTRUCTIONS: [usize; 18] = [
+const NATIVE_INSTRUCTIONS: [usize; 17] = [
     COL_MAP.op.binary_op,
     COL_MAP.op.ternary_op,
     COL_MAP.op.fp254_op,
@@ -24,8 +24,7 @@ const NATIVE_INSTRUCTIONS: [usize; 18] = [
     COL_MAP.op.jumpdest,
     COL_MAP.op.push0,
     // not PUSH (need to increment by more than 1)
-    COL_MAP.op.dup,
-    COL_MAP.op.swap,
+    COL_MAP.op.dup_swap,
     COL_MAP.op.get_context,
     COL_MAP.op.set_context,
     // not EXIT_KERNEL (performs a jump)

evm/src/cpu/decode.rs

@@ -23,7 +23,7 @@ use crate::cpu::columns::{CpuColumnsView, COL_MAP};
 /// behavior.
 /// Note: invalid opcodes are not represented here. _Any_ opcode is permitted to decode to
 /// `is_invalid`. The kernel then verifies that the opcode was _actually_ invalid.
-const OPCODES: [(u8, usize, bool, usize); 17] = [
+const OPCODES: [(u8, usize, bool, usize); 16] = [
     // (start index of block, number of top bits to check (log2), kernel-only, flag column)
     // ADD, MUL, SUB, DIV, MOD, LT, GT and BYTE flags are handled partly manually here, and partly through the Arithmetic table CTL.
     // ADDMOD, MULMOD and SUBMOD flags are handled partly manually here, and partly through the Arithmetic table CTL.
@@ -39,9 +39,8 @@ const OPCODES: [(u8, usize, bool, usize); 17] = [
     (0x58, 0, false, COL_MAP.op.pc),
     (0x5b, 0, false, COL_MAP.op.jumpdest),
     (0x5f, 0, false, COL_MAP.op.push0),
-    (0x60, 5, false, COL_MAP.op.push), // 0x60-0x7f
-    (0x80, 4, false, COL_MAP.op.dup),  // 0x80-0x8f
-    (0x90, 4, false, COL_MAP.op.swap), // 0x90-0x9f
+    (0x60, 5, false, COL_MAP.op.push),     // 0x60-0x7f
+    (0x80, 5, false, COL_MAP.op.dup_swap), // 0x80-0x9f
     (0xee, 0, true, COL_MAP.op.mstore_32bytes),
     (0xf6, 0, true, COL_MAP.op.get_context),
     (0xf7, 0, true, COL_MAP.op.set_context),

evm/src/cpu/dup_swap.rs

@@ -114,7 +114,8 @@ fn eval_packed_dup<P: PackedField>(
     nv: &CpuColumnsView<P>,
     yield_constr: &mut ConstraintConsumer<P>,
 ) {
-    let filter = lv.op.dup;
+    // DUP opcodes have 0 at the 5-th position, while SWAP opcodes have 1.
+    let filter = lv.op.dup_swap * (P::ONES - lv.opcode_bits[4]);
 
     let write_channel = &lv.mem_channels[1];
     let read_channel = &lv.mem_channels[2];
@@ -139,8 +140,10 @@ fn eval_ext_circuit_dup<F: RichField + Extendable<D>, const D: usize>(
     yield_constr: &mut RecursiveConstraintConsumer<F, D>,
 ) {
     let zero = builder.zero_extension();
-
-    let filter = lv.op.dup;
+    let one = builder.one_extension();
+    // DUP opcodes have 0 at the 5-th position, while SWAP opcodes have 1.
+    let mut filter = builder.sub_extension(one, lv.opcode_bits[4]);
+    filter = builder.mul_extension(lv.op.dup_swap, filter);
 
     let write_channel = &lv.mem_channels[1];
     let read_channel = &lv.mem_channels[2];
@@ -187,7 +190,8 @@ fn eval_packed_swap<P: PackedField>(
 ) {
     let n_plus_one = n + P::ONES;
 
-    let filter = lv.op.swap;
+    // DUP opcodes have 0 at the 5-th position, while SWAP opcodes have 1.
+    let filter = lv.op.dup_swap * lv.opcode_bits[4];
 
     let in1_channel = &lv.mem_channels[0];
     let in2_channel = &lv.mem_channels[1];
@@ -215,7 +219,8 @@ fn eval_ext_circuit_swap<F: RichField + Extendable<D>, const D: usize>(
     let one = builder.one_extension();
     let n_plus_one = builder.add_extension(n, one);
 
-    let filter = lv.op.swap;
+    // DUP opcodes have 0 at the 5-th position, while SWAP opcodes have 1.
+    let filter = builder.mul_extension(lv.op.dup_swap, lv.opcode_bits[4]);
 
     let in1_channel = &lv.mem_channels[0];
     let in2_channel = &lv.mem_channels[1];

evm/src/cpu/gas.rs

@@ -34,8 +34,7 @@ const SIMPLE_OPCODES: OpsColumnsView<Option<u32>> = OpsColumnsView {
     jumpdest: G_JUMPDEST,
     push0: G_BASE,
     push: G_VERYLOW,
-    dup: G_VERYLOW,
-    swap: G_VERYLOW,
+    dup_swap: G_VERYLOW,
     get_context: KERNEL_ONLY_INSTR,
     set_context: KERNEL_ONLY_INSTR,
     mstore_32bytes: KERNEL_ONLY_INSTR,

evm/src/cpu/stack.rs

@@ -109,8 +109,7 @@ pub(crate) const STACK_BEHAVIORS: OpsColumnsView<Option<StackBehavior>> = OpsCol
         disable_other_channels: true,
     }),
     push: None, // TODO
-    dup: None,
-    swap: None,
+    dup_swap: None,
     get_context: Some(StackBehavior {
         num_pops: 0,
         pushes: true,

evm/src/witness/transition.rs

@@ -162,8 +162,7 @@ fn fill_op_flag<F: Field>(op: Operation, row: &mut CpuColumnsView<F>) {
     *match op {
         Operation::Push(0) => &mut flags.push0,
         Operation::Push(1..) => &mut flags.push,
-        Operation::Dup(_) => &mut flags.dup,
-        Operation::Swap(_) => &mut flags.swap,
+        Operation::Dup(_) | Operation::Swap(_) => &mut flags.dup_swap,
         Operation::Iszero | Operation::Eq => &mut flags.eq_iszero,
         Operation::Not => &mut flags.not,
         Operation::Syscall(_, _, _) => &mut flags.syscall,
@@ -195,8 +194,7 @@ fn get_op_special_length(op: Operation) -> Option<usize> {
     let behavior_opt = match op {
         Operation::Push(0) => STACK_BEHAVIORS.push0,
         Operation::Push(1..) => STACK_BEHAVIORS.push,
-        Operation::Dup(_) => STACK_BEHAVIORS.dup,
-        Operation::Swap(_) => STACK_BEHAVIORS.swap,
+        Operation::Dup(_) | Operation::Swap(_) => STACK_BEHAVIORS.dup_swap,
         Operation::Iszero => IS_ZERO_STACK_BEHAVIOR,
         Operation::Not => STACK_BEHAVIORS.not,
         Operation::Syscall(_, _, _) => STACK_BEHAVIORS.syscall,