scroll-dev-1019: squash all old diffs

crates/core/executor/Cargo.toml

@@ -28,6 +28,7 @@ bincode = "1.3.3"
 hashbrown = { version = "0.14.5", features = ["serde", "inline-more"] }
 itertools = "0.13.0"
 rand = "0.8.5"
+generic-array = { version = "1.1.0", features = ["alloc", "serde"] }
 num = { version = "0.4.3" }
 typenum = "1.17.0"
 nohash-hasher = "0.2.0"

crates/core/executor/src/events/memcpy.rs

@@ -0,0 +1,15 @@
+use super::{LookupId, MemoryLocalEvent, MemoryReadRecord, MemoryWriteRecord};
+use serde::{Deserialize, Serialize};
+
+#[derive(Default, Clone, Debug, Serialize, Deserialize)]
+pub struct MemCopyEvent {
+    pub lookup_id: LookupId,
+    pub shard: u32,
+    pub clk: u32,
+    pub src_ptr: u32,
+    pub dst_ptr: u32,
+    pub read_records: Vec<MemoryReadRecord>,
+    pub write_records: Vec<MemoryWriteRecord>,
+    /// The local memory access records.
+    pub local_mem_access: Vec<MemoryLocalEvent>,
+}

crates/core/executor/src/events/mod.rs

@@ -3,6 +3,7 @@
 mod alu;
 mod byte;
 mod cpu;
+mod memcpy;
 mod memory;
 mod precompiles;
 mod syscall;
@@ -11,6 +12,7 @@ mod utils;
 pub use alu::*;
 pub use byte::*;
 pub use cpu::*;
+pub use memcpy::*;
 pub use memory::*;
 pub use precompiles::*;
 pub use syscall::*;

crates/core/executor/src/events/precompiles/bn254_scalar.rs

@@ -0,0 +1,166 @@
+use num::BigUint;
+use sp1_curves::{
+    params::{FieldParameters, NumWords},
+    weierstrass::bn254::Bn254ScalarField,
+};
+use typenum::Unsigned;
+
+use serde::{Deserialize, Serialize};
+
+use crate::{
+    events::{LookupId, MemoryLocalEvent, MemoryReadRecord, MemoryWriteRecord},
+    syscalls::SyscallContext,
+};
+
+use super::FieldOperation;
+
+pub const NUM_WORDS_PER_FE: usize = 8;
+
+#[derive(Default, PartialEq, Copy, Clone, Debug, Serialize, Deserialize)]
+pub enum Bn254FieldOperation {
+    #[default]
+    Invalid = 0,
+    Mul = 2,
+    Mac = 4,
+}
+
+impl Bn254FieldOperation {
+    pub const fn to_field_operation(&self) -> FieldOperation {
+        match self {
+            Bn254FieldOperation::Mul => FieldOperation::Mul,
+            Bn254FieldOperation::Mac => panic!("not supported"),
+            Bn254FieldOperation::Invalid => panic!("what??"),
+        }
+    }
+}
+
+#[derive(Default, Clone, Debug, Serialize, Deserialize)]
+pub struct Bn254FieldArithEvent {
+    pub lookup_id: LookupId,
+    pub shard: u32,
+    pub clk: u32,
+    pub op: Bn254FieldOperation,
+    pub arg1: FieldArithMemoryAccess<MemoryWriteRecord>,
+    pub arg2: FieldArithMemoryAccess<MemoryReadRecord>,
+    pub a: Option<FieldArithMemoryAccess<MemoryReadRecord>>,
+    pub b: Option<FieldArithMemoryAccess<MemoryReadRecord>>,
+    /// The local memory access records.
+    pub local_mem_access: Vec<MemoryLocalEvent>,
+}
+
+pub fn create_bn254_scalar_arith_event(
+    rt: &mut SyscallContext,
+    arg1: u32,
+    arg2: u32,
+    op: Bn254FieldOperation,
+) -> Bn254FieldArithEvent {
+    let start_clk = rt.clk;
+    let p_ptr = arg1;
+    let q_ptr = arg2;
+
+    assert_eq!(p_ptr % 4, 0, "p_ptr({p_ptr:x}) is not aligned");
+    assert_eq!(q_ptr % 4, 0, "q_ptr({q_ptr:x}) is not aligned");
+
+    let nw_per_fe = <Bn254ScalarField as NumWords>::WordsFieldElement::USIZE;
+    debug_assert_eq!(nw_per_fe, NUM_WORDS_PER_FE);
+
+    let arg1: Vec<u32> = rt.slice_unsafe(p_ptr, nw_per_fe);
+    let arg2 = match op {
+        // 2 ptrs of real U256 values
+        Bn254FieldOperation::Mac => FieldArithMemoryAccess::read(rt, arg2, 2),
+        _ => FieldArithMemoryAccess::read(rt, arg2, nw_per_fe),
+    };
+
+    let bn_arg1 = BigUint::from_bytes_le(
+        &arg1.iter().copied().flat_map(u32::to_le_bytes).collect::<Vec<u8>>(),
+    );
+    let modulus = Bn254ScalarField::modulus();
+
+    let (a, b, bn_arg1_out) = if matches!(op, Bn254FieldOperation::Mac) {
+        let a = FieldArithMemoryAccess::read(rt, arg2.memory_records[0].value, nw_per_fe);
+        let b = FieldArithMemoryAccess::read(rt, arg2.memory_records[1].value, nw_per_fe);
+
+        let bn_a = a.value_as_biguint();
+        let bn_b = b.value_as_biguint();
+        let bn_arg1_out = (&bn_a * &bn_b + &bn_arg1) % modulus;
+
+        (Some(a), Some(b), bn_arg1_out)
+    } else {
+        let bn_arg2 = arg2.value_as_biguint();
+
+        let bn_arg1_out = match op {
+            Bn254FieldOperation::Mul => (&bn_arg1 * &bn_arg2) % modulus,
+            _ => unimplemented!("not supported"),
+        };
+        (None, None, bn_arg1_out)
+    };
+
+    log::trace!(
+        "shard: {}, clk: {}, op: {:?}, arg1: {:?}, arg2: {:?}, a: {:?}, b: {:?}",
+        rt.current_shard(),
+        rt.clk,
+        op,
+        arg1,
+        arg2,
+        a,
+        b
+    );
+    rt.clk += 1;
+
+    let mut result_words = bn_arg1_out.to_u32_digits();
+    result_words.resize(nw_per_fe, 0);
+
+    let arg1 = FieldArithMemoryAccess::write(rt, p_ptr, &result_words);
+
+    let shard = rt.current_shard();
+    Bn254FieldArithEvent {
+        lookup_id: rt.syscall_lookup_id,
+        shard,
+        clk: start_clk,
+        op,
+        arg1,
+        arg2,
+        a,
+        b,
+        local_mem_access: rt.postprocess(),
+    }
+}
+
+#[derive(Default, Clone, Debug, Serialize, Deserialize)]
+pub struct FieldArithMemoryAccess<T> {
+    pub ptr: u32,
+    pub memory_records: Vec<T>,
+}
+
+impl FieldArithMemoryAccess<MemoryReadRecord> {
+    pub fn read(rt: &mut SyscallContext, ptr: u32, len: usize) -> Self {
+        let (memory_records, _) = rt.mr_slice(ptr, len);
+        Self { ptr, memory_records }
+    }
+
+    pub fn value_as_biguint(&self) -> BigUint {
+        BigUint::from_bytes_le(
+            &self
+                .memory_records
+                .iter()
+                .flat_map(|word| word.value.to_le_bytes())
+                .collect::<Vec<u8>>(),
+        )
+    }
+}
+
+impl FieldArithMemoryAccess<MemoryWriteRecord> {
+    pub fn write(rt: &mut SyscallContext, ptr: u32, values: &[u32]) -> Self {
+        Self { ptr, memory_records: rt.mw_slice(ptr, values) }
+    }
+
+    pub fn prev_value_as_biguint(&self) -> BigUint {
+        BigUint::from_bytes_le(
+            &self
+                .memory_records
+                .iter()
+                .flat_map(|word| word.prev_value.to_le_bytes())
+                .collect::<Vec<u8>>(),
+        )
+    }
+}

crates/core/executor/src/events/precompiles/mod.rs

@@ -1,3 +1,4 @@
+mod bn254_scalar;
 mod ec;
 mod edwards;
 mod fptower;
@@ -6,6 +7,9 @@ mod sha256_compress;
 mod sha256_extend;
 mod uint256;
 
+pub use bn254_scalar::{
+    create_bn254_scalar_arith_event, Bn254FieldArithEvent, Bn254FieldOperation, NUM_WORDS_PER_FE,
+};
 pub use ec::*;
 pub use edwards::*;
 pub use fptower::*;
@@ -19,6 +23,7 @@ pub use uint256::*;
 
 use crate::syscalls::SyscallCode;
 
+use super::{MemCopyEvent};
 use super::{MemoryLocalEvent, SyscallEvent};
 
 #[derive(Clone, Debug, Serialize, Deserialize, EnumIter)]
@@ -52,6 +57,12 @@ pub enum PrecompileEvent {
     Bn254Fp2AddSub(Fp2AddSubEvent),
     /// Bn254 quadratic field mul precompile event.
     Bn254Fp2Mul(Fp2MulEvent),
+
+    Bn254ScalarMac(Bn254FieldArithEvent),
+    Bn254ScalarMul(Bn254FieldArithEvent),
+    MemCopy32(MemCopyEvent),
+    MemCopy64(MemCopyEvent),
+
     /// Bls12-381 curve add precompile event.
     Bls12381Add(EllipticCurveAddEvent),
     /// Bls12-381 curve double precompile event.
@@ -120,6 +131,12 @@ impl PrecompileLocalMemory for Vec<(SyscallEvent, PrecompileEvent)> {
                 PrecompileEvent::Bls12381Fp2Mul(e) | PrecompileEvent::Bn254Fp2Mul(e) => {
                     iterators.push(e.local_mem_access.iter());
                 }
+                PrecompileEvent::Bn254ScalarMac(e) | PrecompileEvent::Bn254ScalarMul(e) => {
+                    iterators.push(e.local_mem_access.iter());
+                }
+                PrecompileEvent::MemCopy32(e) | PrecompileEvent::MemCopy64(e) => {
+                    iterators.push(e.local_mem_access.iter());
+                }
             }
         }
 

crates/core/executor/src/executor.rs

@@ -899,6 +899,10 @@ impl<'a> Executor<'a> {
                 let value = (memory_read_value).to_le_bytes()[(addr % 4) as usize];
                 a = ((value as i8) as i32) as u32;
                 memory_store_value = Some(memory_read_value);
+                //println!(
+                //    "[clk: {}, pc: 0x{:x}] LB: {:?} <- {:x}",
+                //    self.state.global_clk, self.state.pc, rd, a
+                //);
                 self.rw(rd, a);
             }
             Opcode::LH => {
@@ -913,6 +917,10 @@ impl<'a> Executor<'a> {
                 };
                 a = ((value as i16) as i32) as u32;
                 memory_store_value = Some(memory_read_value);
+                //println!(
+                //    "[clk: {}, pc: 0x{:x}] LH: {:?} <- {:x}",
+                //    self.state.global_clk, self.state.pc, rd, a
+                //);
                 self.rw(rd, a);
             }
             Opcode::LW => {
@@ -922,6 +930,10 @@ impl<'a> Executor<'a> {
                 }
                 a = memory_read_value;
                 memory_store_value = Some(memory_read_value);
+                //println!(
+                //    "[clk: {}, pc: 0x{:x}] LW: {:?} <- {}",
+                //    self.state.global_clk, self.state.pc, rd, a
+                //);
                 self.rw(rd, a);
             }
             Opcode::LBU => {
@@ -957,6 +969,10 @@ impl<'a> Executor<'a> {
                     _ => unreachable!(),
                 };
                 memory_store_value = Some(value);
+                //println!(
+                //    "[clk: {}, pc: 0x{:x}] SB 0x{:x} <- 0x{:x}",
+                //    self.state.global_clk, pc, addr, value
+                //);
                 self.mw_cpu(align(addr), value, MemoryAccessPosition::Memory);
             }
             Opcode::SH => {
@@ -970,6 +986,10 @@ impl<'a> Executor<'a> {
                     _ => unreachable!(),
                 };
                 memory_store_value = Some(value);
+                //println!(
+                //    "[clk: {}, pc: 0x{:x}] SH 0x{:x} <- 0x{:x}",
+                //    self.state.global_clk, pc, addr, value
+                //);
                 self.mw_cpu(align(addr), value, MemoryAccessPosition::Memory);
             }
             Opcode::SW => {
@@ -979,6 +999,10 @@ impl<'a> Executor<'a> {
                 }
                 let value = a;
                 memory_store_value = Some(value);
+                //println!(
+                //    "[clk: {}, pc: 0x{:x}] SW 0x{:x} <- 0x{:x}",
+                //    self.state.global_clk, pc, addr, value
+                //);
                 self.mw_cpu(align(addr), value, MemoryAccessPosition::Memory);
             }
 
@@ -1070,6 +1094,7 @@ impl<'a> Executor<'a> {
                     return Err(ExecutionError::InvalidSyscallUsage(syscall_id as u64));
                 }
 
+                let global_clk = self.state.global_clk;
                 // Update the syscall counts.
                 let syscall_for_count = syscall.count_map();
                 let syscall_count = self.state.syscall_counts.entry(syscall_for_count).or_insert(0);
@@ -1089,6 +1114,14 @@ impl<'a> Executor<'a> {
                 }
                 let mut precompile_rt = SyscallContext::new(self);
                 precompile_rt.syscall_lookup_id = syscall_lookup_id;
+                log::trace!(
+                    "[clk: {}, pc: 0x{:x}] ecall syscall_id=0x{:x}, b: 0x{:x}, c: 0x{:x}",
+                    global_clk,
+                    pc,
+                    syscall_id,
+                    b,
+                    c,
+                );
                 let (precompile_next_pc, precompile_cycles, returned_exit_code) =
                     if let Some(syscall_impl) = syscall_impl {
                         // Executing a syscall optionally returns a value to write to the t0
@@ -1125,6 +1158,10 @@ impl<'a> Executor<'a> {
                 next_pc = precompile_next_pc;
                 self.state.clk += precompile_cycles;
                 exit_code = returned_exit_code;
+
+                //log::info!(
+                //    "execute_instruction {syscall:?} {syscall_count} {nonce} {syscall_lookup_id}"
+                //);
             }
             Opcode::EBREAK => {
                 return Err(ExecutionError::Breakpoint());

crates/core/executor/src/syscalls/code.rs

@@ -128,6 +128,15 @@ pub enum SyscallCode {
 
     /// Executes the `BN254_FP2_MUL` precompile.
     BN254_FP2_MUL = 0x00_01_01_2B,
+
+    /// Execute the `MEMCPY_32` precompile.
+    MEMCPY_32 = 0x00_01_01_90,
+    /// Execute the `MEMCPY_64` precompile.
+    MEMCPY_64 = 0x00_01_01_91,
+    /// Execute the `BN254_SCALAR_MUL` precompile.
+    BN254_SCALAR_MUL = 0x00_01_01_80,
+    /// Execute the `BN254_SCALAR_MAC` precompile.
+    BN254_SCALAR_MAC = 0x00_01_01_81,
 }
 
 impl SyscallCode {
@@ -170,6 +179,10 @@ impl SyscallCode {
             0x00_01_01_2A => SyscallCode::BN254_FP2_SUB,
             0x00_01_01_2B => SyscallCode::BN254_FP2_MUL,
             0x00_00_01_1C => SyscallCode::BLS12381_DECOMPRESS,
+            0x00_01_01_90 => SyscallCode::MEMCPY_32,
+            0x00_01_01_91 => SyscallCode::MEMCPY_64,
+            0x00_01_01_80 => SyscallCode::BN254_SCALAR_MUL,
+            0x00_01_01_81 => SyscallCode::BN254_SCALAR_MAC,
             _ => panic!("invalid syscall number: {value}"),
         }
     }