Memory channel for program counter (#717)

evm/src/all_stark.rs

@@ -1,3 +1,5 @@
+use std::iter;
+
 use plonky2::field::extension::Extendable;
 use plonky2::field::types::Field;
 use plonky2::hash::hash_types::RichField;
@@ -5,6 +7,7 @@ use plonky2::hash::hash_types::RichField;
 use crate::config::StarkConfig;
 use crate::cpu::cpu_stark;
 use crate::cpu::cpu_stark::CpuStark;
+use crate::cpu::membus::NUM_GP_CHANNELS;
 use crate::cross_table_lookup::{CrossTableLookup, TableWithColumns};
 use crate::keccak::keccak_stark;
 use crate::keccak::keccak_stark::KeccakStark;
@@ -13,8 +16,8 @@ use crate::keccak_memory::keccak_memory_stark;
 use crate::keccak_memory::keccak_memory_stark::KeccakMemoryStark;
 use crate::logic;
 use crate::logic::LogicStark;
+use crate::memory::memory_stark;
 use crate::memory::memory_stark::MemoryStark;
-use crate::memory::{memory_stark, NUM_CHANNELS};
 use crate::stark::Stark;
 
 #[derive(Clone)]
@@ -129,11 +132,16 @@ fn ctl_logic<F: Field>() -> CrossTableLookup<F> {
 }
 
 fn ctl_memory<F: Field>() -> CrossTableLookup<F> {
-    let cpu_memory_ops = (0..NUM_CHANNELS).map(|channel| {
+    let cpu_memory_code_read = TableWithColumns::new(
+        Table::Cpu,
+        cpu_stark::ctl_data_code_memory(),
+        Some(cpu_stark::ctl_filter_code_memory()),
+    );
+    let cpu_memory_gp_ops = (0..NUM_GP_CHANNELS).map(|channel| {
         TableWithColumns::new(
             Table::Cpu,
-            cpu_stark::ctl_data_memory(channel),
+            cpu_stark::ctl_data_gp_memory(channel),
-            Some(cpu_stark::ctl_filter_memory(channel)),
+            Some(cpu_stark::ctl_filter_gp_memory(channel)),
         )
     });
     let keccak_memory_reads = (0..KECCAK_WIDTH_BYTES).map(|i| {
@@ -150,7 +158,8 @@ fn ctl_memory<F: Field>() -> CrossTableLookup<F> {
             Some(keccak_memory_stark::ctl_filter()),
         )
     });
-    let all_lookers = cpu_memory_ops
+    let all_lookers = iter::once(cpu_memory_code_read)
+        .chain(cpu_memory_gp_ops)
         .chain(keccak_memory_reads)
         .chain(keccak_memory_writes)
         .collect();
@@ -725,6 +734,7 @@ mod tests {
     }
 
     #[test]
+    #[ignore] // Ignoring but not deleting so the test can serve as an API usage example
     fn test_all_stark() -> Result<()> {
         let config = StarkConfig::standard_fast_config();
         let (all_stark, proof) = get_proof(&config)?;
@@ -732,6 +742,7 @@ mod tests {
     }
 
     #[test]
+    #[ignore] // Ignoring but not deleting so the test can serve as an API usage example
     fn test_all_stark_recursive_verifier() -> Result<()> {
         init_logger();
 

evm/src/cpu/columns/mod.rs

@@ -7,6 +7,7 @@ use std::mem::{size_of, transmute};
 use std::ops::{Index, IndexMut};
 
 use crate::cpu::columns::general::CpuGeneralColumnsView;
+use crate::cpu::membus::NUM_GP_CHANNELS;
 use crate::memory;
 use crate::util::{indices_arr, transmute_no_compile_time_size_checks};
 
@@ -35,6 +36,13 @@ pub struct CpuColumnsView<T: Copy> {
     /// Lets us re-use columns in non-cycle rows.
     pub is_cpu_cycle: T,
 
+    /// If CPU cycle: Current context.
+    // TODO: this is currently unconstrained
+    pub context: T,
+
+    /// If CPU cycle: Context for code memory channel.
+    pub code_context: T,
+
     /// If CPU cycle: The program counter for the current instruction.
     pub program_counter: T,
 
@@ -159,7 +167,7 @@ pub struct CpuColumnsView<T: Copy> {
     pub(crate) general: CpuGeneralColumnsView<T>,
 
     pub(crate) clock: T,
-    pub mem_channels: [MemoryChannelView<T>; memory::NUM_CHANNELS],
+    pub mem_channels: [MemoryChannelView<T>; NUM_GP_CHANNELS],
 }
 
 // `u8` is guaranteed to have a `size_of` of 1.

evm/src/cpu/cpu_stark.rs

@@ -1,4 +1,5 @@
 use std::borrow::{Borrow, BorrowMut};
+use std::iter::repeat;
 use std::marker::PhantomData;
 
 use itertools::Itertools;
@@ -10,10 +11,11 @@ use plonky2::hash::hash_types::RichField;
 use crate::constraint_consumer::{ConstraintConsumer, RecursiveConstraintConsumer};
 use crate::cpu::columns::{CpuColumnsView, COL_MAP, NUM_CPU_COLUMNS};
 use crate::cpu::{
-    bootstrap_kernel, control_flow, decode, jumps, simple_logic, stack_bounds, syscalls,
+    bootstrap_kernel, control_flow, decode, jumps, membus, simple_logic, stack_bounds, syscalls,
 };
 use crate::cross_table_lookup::Column;
-use crate::memory::NUM_CHANNELS;
+use crate::memory::segments::Segment;
+use crate::memory::{NUM_CHANNELS, VALUE_LIMBS};
 use crate::stark::Stark;
 use crate::vars::{StarkEvaluationTargets, StarkEvaluationVars};
 
@@ -25,14 +27,13 @@ pub fn ctl_data_keccak<F: Field>() -> Vec<Column<F>> {
 }
 
 pub fn ctl_data_keccak_memory<F: Field>() -> Vec<Column<F>> {
-    // When executing KECCAK_GENERAL, the memory channels are used as follows:
+    // When executing KECCAK_GENERAL, the GP memory channels are used as follows:
-    // channel 0: instruction
+    // GP channel 0: stack[-1] = context
-    // channel 1: stack[-1] = context
+    // GP channel 1: stack[-2] = segment
-    // channel 2: stack[-2] = segment
+    // GP channel 2: stack[-3] = virtual
-    // channel 3: stack[-3] = virtual
+    let context = Column::single(COL_MAP.mem_channels[0].value[0]);
-    let context = Column::single(COL_MAP.mem_channels[1].value[0]);
+    let segment = Column::single(COL_MAP.mem_channels[1].value[0]);
-    let segment = Column::single(COL_MAP.mem_channels[2].value[0]);
+    let virt = Column::single(COL_MAP.mem_channels[2].value[0]);
-    let virt = Column::single(COL_MAP.mem_channels[3].value[0]);
 
     let num_channels = F::from_canonical_usize(NUM_CHANNELS);
     let clock = Column::linear_combination([(COL_MAP.clock, num_channels)]);
@@ -60,29 +61,57 @@ pub fn ctl_filter_logic<F: Field>() -> Column<F> {
     Column::sum([COL_MAP.is_and, COL_MAP.is_or, COL_MAP.is_xor])
 }
 
-pub fn ctl_data_memory<F: Field>(channel: usize) -> Vec<Column<F>> {
+pub const MEM_CODE_CHANNEL_IDX: usize = 0;
-    debug_assert!(channel < NUM_CHANNELS);
+pub const MEM_GP_CHANNELS_IDX_START: usize = MEM_CODE_CHANNEL_IDX + 1;
+
+/// Make the time/channel column for memory lookups.
+fn mem_time_and_channel<F: Field>(channel: usize) -> Column<F> {
+    let scalar = F::from_canonical_usize(NUM_CHANNELS);
+    let addend = F::from_canonical_usize(channel);
+    Column::linear_combination_with_constant([(COL_MAP.clock, scalar)], addend)
+}
+
+pub fn ctl_data_code_memory<F: Field>() -> Vec<Column<F>> {
+    let mut cols = vec![
+        Column::constant(F::ONE),                                      // is_read
+        Column::single(COL_MAP.code_context),                          // addr_context
+        Column::constant(F::from_canonical_u64(Segment::Code as u64)), // addr_segment
+        Column::single(COL_MAP.program_counter),                       // addr_virtual
+    ];
+
+    // Low limb of the value matches the opcode bits
+    cols.push(Column::le_bits(COL_MAP.opcode_bits));
+
+    // High limbs of the value are all zero.
+    cols.extend(repeat(Column::constant(F::ZERO)).take(VALUE_LIMBS - 1));
+
+    cols.push(mem_time_and_channel(MEM_CODE_CHANNEL_IDX));
+
+    cols
+}
+
+pub fn ctl_data_gp_memory<F: Field>(channel: usize) -> Vec<Column<F>> {
     let channel_map = COL_MAP.mem_channels[channel];
-    let mut cols: Vec<Column<F>> = Column::singles([
+    let mut cols: Vec<_> = Column::singles([
         channel_map.is_read,
         channel_map.addr_context,
         channel_map.addr_segment,
         channel_map.addr_virtual,
     ])
-    .collect_vec();
+    .collect();
+
     cols.extend(Column::singles(channel_map.value));
 
-    let scalar = F::from_canonical_usize(NUM_CHANNELS);
+    cols.push(mem_time_and_channel(MEM_GP_CHANNELS_IDX_START + channel));
-    let addend = F::from_canonical_usize(channel);
-    cols.push(Column::linear_combination_with_constant(
-        [(COL_MAP.clock, scalar)],
-        addend,
-    ));
 
     cols
 }
 
-pub fn ctl_filter_memory<F: Field>(channel: usize) -> Column<F> {
+pub fn ctl_filter_code_memory<F: Field>() -> Column<F> {
+    Column::single(COL_MAP.is_cpu_cycle)
+}
+
+pub fn ctl_filter_gp_memory<F: Field>(channel: usize) -> Column<F> {
     Column::single(COL_MAP.mem_channels[channel].used)
 }
 
@@ -95,6 +124,7 @@ impl<F: RichField, const D: usize> CpuStark<F, D> {
     pub fn generate(&self, local_values: &mut [F; NUM_CPU_COLUMNS]) {
         let local_values: &mut CpuColumnsView<_> = local_values.borrow_mut();
         decode::generate(local_values);
+        membus::generate(local_values);
         simple_logic::generate(local_values);
         stack_bounds::generate(local_values); // Must come after `decode`.
     }
@@ -117,6 +147,7 @@ impl<F: RichField + Extendable<D>, const D: usize> Stark<F, D> for CpuStark<F, D
         control_flow::eval_packed_generic(local_values, next_values, yield_constr);
         decode::eval_packed_generic(local_values, yield_constr);
         jumps::eval_packed(local_values, next_values, yield_constr);
+        membus::eval_packed(local_values, yield_constr);
         simple_logic::eval_packed(local_values, yield_constr);
         stack_bounds::eval_packed(local_values, yield_constr);
         syscalls::eval_packed(local_values, next_values, yield_constr);
@@ -134,6 +165,7 @@ impl<F: RichField + Extendable<D>, const D: usize> Stark<F, D> for CpuStark<F, D
         control_flow::eval_ext_circuit(builder, local_values, next_values, yield_constr);
         decode::eval_ext_circuit(builder, local_values, yield_constr);
         jumps::eval_ext_circuit(builder, local_values, next_values, yield_constr);
+        membus::eval_ext_circuit(builder, local_values, yield_constr);
         simple_logic::eval_ext_circuit(builder, local_values, yield_constr);
         stack_bounds::eval_ext_circuit(builder, local_values, yield_constr);
         syscalls::eval_ext_circuit(builder, local_values, next_values, yield_constr);

evm/src/cpu/membus.rs

@@ -0,0 +1,70 @@
+use plonky2::field::extension::Extendable;
+use plonky2::field::packed::PackedField;
+use plonky2::field::types::PrimeField64;
+use plonky2::hash::hash_types::RichField;
+use plonky2::iop::ext_target::ExtensionTarget;
+
+use crate::constraint_consumer::{ConstraintConsumer, RecursiveConstraintConsumer};
+use crate::cpu::columns::CpuColumnsView;
+
+/// General-purpose memory channels; they can read and write to all contexts/segments/addresses.
+pub const NUM_GP_CHANNELS: usize = 4;
+
+pub mod channel_indices {
+    use std::ops::Range;
+
+    pub const CODE: usize = 0;
+    pub const GP: Range<usize> = CODE + 1..(CODE + 1) + super::NUM_GP_CHANNELS;
+}
+
+/// Total memory channels used by the CPU table. This includes all the `GP_MEM_CHANNELS` as well as
+/// all special-purpose memory channels.
+///
+/// Currently, there is one special-purpose memory channel, which reads the opcode from memory. Its
+/// limitations are:
+///  - it is enabled by `is_cpu_cycle`,
+///  - it always reads and cannot write,
+///  - the context is derived from the current context and the `is_kernel_mode` flag,
+///  - the segment is hard-wired to the code segment,
+///  - the address is `program_counter`,
+///  - the value must fit in one byte (in the least-significant position) and its eight bits are
+///    found in `opcode_bits`.
+/// These limitations save us numerous columns in the CPU table.
+pub const NUM_CHANNELS: usize = channel_indices::GP.end;
+
+/// Calculates `lv.stack_len_bounds_aux`. Note that this must be run after decode.
+pub fn generate<F: PrimeField64>(lv: &mut CpuColumnsView<F>) {
+    let cycle_filter = lv.is_cpu_cycle;
+    if cycle_filter == F::ZERO {
+        return;
+    }
+
+    assert!(lv.is_kernel_mode.to_canonical_u64() <= 1);
+
+    // Set `lv.code_context` to 0 if in kernel mode and to `lv.context` if in user mode.
+    lv.code_context = (F::ONE - lv.is_kernel_mode) * lv.context;
+}
+
+pub fn eval_packed<P: PackedField>(
+    lv: &CpuColumnsView<P>,
+    yield_constr: &mut ConstraintConsumer<P>,
+) {
+    // Validate `lv.code_context`. It should be 0 if in kernel mode and `lv.context` if in user
+    // mode.
+    yield_constr.constraint(
+        lv.is_cpu_cycle * (lv.code_context - (P::ONES - lv.is_kernel_mode) * lv.context),
+    );
+}
+
+pub fn eval_ext_circuit<F: RichField + Extendable<D>, const D: usize>(
+    builder: &mut plonky2::plonk::circuit_builder::CircuitBuilder<F, D>,
+    lv: &CpuColumnsView<ExtensionTarget<D>>,
+    yield_constr: &mut RecursiveConstraintConsumer<F, D>,
+) {
+    // Validate `lv.code_context`. It should be 0 if in kernel mode and `lv.context` if in user
+    // mode.
+    let diff = builder.sub_extension(lv.context, lv.code_context);
+    let constr = builder.mul_sub_extension(lv.is_kernel_mode, lv.context, diff);
+    let filtered_constr = builder.mul_extension(lv.is_cpu_cycle, constr);
+    yield_constr.constraint(builder, filtered_constr);
+}

evm/src/cpu/mod.rs

@@ -5,6 +5,7 @@ pub mod cpu_stark;
 pub(crate) mod decode;
 mod jumps;
 pub mod kernel;
+pub(crate) mod membus;
 mod simple_logic;
 mod stack_bounds;
 mod syscalls;

evm/src/memory/mod.rs

@@ -3,5 +3,5 @@ pub mod memory_stark;
 pub mod segments;
 
 // TODO: Move to CPU module, now that channels have been removed from the memory table.
-pub(crate) const NUM_CHANNELS: usize = 4;
+pub(crate) const NUM_CHANNELS: usize = crate::cpu::membus::NUM_CHANNELS;
 pub(crate) const VALUE_LIMBS: usize = 8;