From 0ed3e1d29dca7d27e0d3aa1780f7cacc9481aa46 Mon Sep 17 00:00:00 2001 From: Henry Date: Tue, 31 Mar 2026 23:43:20 +0200 Subject: chore: add more superinstructions Signed-off-by: Henry --- crates/parser/src/visit.rs | 131 ++++++++++++++++++--- crates/tinywasm/benches/argon2id.rs | 8 +- crates/tinywasm/benches/tinywasm.rs | 8 +- crates/tinywasm/src/interpreter/executor.rs | 102 ++++++++-------- .../tinywasm/src/interpreter/stack/value_stack.rs | 38 +----- crates/tinywasm/src/interpreter/values.rs | 37 ++++-- crates/types/src/instructions.rs | 7 +- 7 files changed, 210 insertions(+), 121 deletions(-) (limited to 'crates') diff --git a/crates/parser/src/visit.rs b/crates/parser/src/visit.rs index d02eb11..b71185e 100644 --- a/crates/parser/src/visit.rs +++ b/crates/parser/src/visit.rs @@ -325,11 +325,7 @@ impl FunctionBuilder { }; match frame.kind { - FrameKind::Loop => { - if let Instruction::Jump(target) = &mut self.instructions[jump_ip] { - *target = self.ctx_stack[ctx_idx].start_ip as u32; - } - } + FrameKind::Loop => self.patch_jump(jump_ip, self.ctx_stack[ctx_idx].start_ip), _ => self.ctx_stack[ctx_idx].branch_jumps.push(jump_ip), } } @@ -385,12 +381,12 @@ impl<'a, R: WasmModuleResources> wasmparser::VisitOperator<'a> for FunctionBuild } define_mem_operands! { - visit_i32_load(I32Load), visit_i64_load(I64Load), visit_f32_load(F32Load), visit_f64_load(F64Load), visit_i32_load8_s(I32Load8S), visit_i32_load8_u(I32Load8U), visit_i32_load16_s(I32Load16S), visit_i32_load16_u(I32Load16U), visit_i64_load8_s(I64Load8S), visit_i64_load8_u(I64Load8U), visit_i64_load16_s(I64Load16S), visit_i64_load16_u(I64Load16U), visit_i64_load32_s(I64Load32S), visit_i64_load32_u(I64Load32U), visit_i32_store( I32Store), visit_i64_store(I64Store), visit_f32_store(F32Store), visit_f64_store(F64Store), visit_i32_store8(I32Store8), visit_i32_store16(I32Store16), visit_i64_store8(I64Store8), visit_i64_store16(I64Store16), visit_i64_store32(I64Store32) + visit_i32_load(I32Load), visit_i64_load(I64Load), visit_f32_load(F32Load), visit_f64_load(F64Load), visit_i32_load8_s(I32Load8S), visit_i32_load8_u(I32Load8U), visit_i32_load16_s(I32Load16S), visit_i32_load16_u(I32Load16U), visit_i64_load8_s(I64Load8S), visit_i64_load8_u(I64Load8U), visit_i64_load16_s(I64Load16S), visit_i64_load16_u(I64Load16U), visit_i64_load32_s(I64Load32S), visit_i64_load32_u(I64Load32U), visit_i64_store(I64Store), visit_f32_store(F32Store), visit_f64_store(F64Store), visit_i32_store8(I32Store8), visit_i32_store16(I32Store16), visit_i64_store8(I64Store8), visit_i64_store16(I64Store16), visit_i64_store32(I64Store32) } define_operands! { // basic instructions - visit_global_get(GlobalGet, u32), visit_i32_const(I32Const, i32), visit_i64_const(I64Const, i64), visit_call(Call, u32), visit_return_call(ReturnCall, u32), visit_memory_size(MemorySize, u32), visit_memory_grow(MemoryGrow, u32), visit_unreachable(Unreachable), visit_nop(Nop), visit_i32_eqz(I32Eqz), visit_i32_eq(I32Eq), visit_i32_ne(I32Ne), visit_i32_lt_s(I32LtS), visit_i32_lt_u(I32LtU), visit_i32_gt_s(I32GtS), visit_i32_gt_u(I32GtU), visit_i32_le_s(I32LeS), visit_i32_le_u(I32LeU), visit_i32_ge_s(I32GeS), visit_i32_ge_u(I32GeU), visit_i64_eqz(I64Eqz), visit_i64_eq(I64Eq), visit_i64_ne(I64Ne), visit_i64_lt_s(I64LtS), visit_i64_lt_u(I64LtU), visit_i64_gt_s(I64GtS), visit_i64_gt_u(I64GtU), visit_i64_le_s(I64LeS), visit_i64_le_u(I64LeU), visit_i64_ge_s(I64GeS), visit_i64_ge_u(I64GeU), visit_f32_eq(F32Eq), visit_f32_ne(F32Ne), visit_f32_lt(F32Lt), visit_f32_gt(F32Gt), visit_f32_le(F32Le), visit_f32_ge(F32Ge), visit_f64_eq(F64Eq), visit_f64_ne(F64Ne), visit_f64_lt(F64Lt), visit_f64_gt(F64Gt), visit_f64_le(F64Le), visit_f64_ge(F64Ge), visit_i32_clz(I32Clz), visit_i32_ctz(I32Ctz), visit_i32_popcnt(I32Popcnt), visit_i32_add(I32Add), visit_i32_sub(I32Sub), visit_i32_mul(I32Mul), visit_i32_div_s(I32DivS), visit_i32_div_u(I32DivU), visit_i32_rem_s(I32RemS), visit_i32_rem_u(I32RemU), visit_i32_and(I32And), visit_i32_or(I32Or), visit_i32_xor(I32Xor), visit_i32_shl(I32Shl), visit_i32_shr_s(I32ShrS), visit_i32_shr_u(I32ShrU), visit_i32_rotl(I32Rotl), visit_i32_rotr(I32Rotr), visit_i64_clz(I64Clz), visit_i64_ctz(I64Ctz), visit_i64_popcnt(I64Popcnt), visit_i64_add(I64Add), visit_i64_sub(I64Sub), visit_i64_mul(I64Mul), visit_i64_div_s(I64DivS), visit_i64_div_u(I64DivU), visit_i64_rem_s(I64RemS), visit_i64_rem_u(I64RemU), visit_i64_and(I64And), visit_i64_or(I64Or), visit_i64_xor(I64Xor), visit_i64_shl(I64Shl), visit_i64_shr_s(I64ShrS), visit_i64_shr_u(I64ShrU), visit_i64_rotl(I64Rotl), visit_i64_rotr(I64Rotr), visit_f32_abs(F32Abs), visit_f32_neg(F32Neg), visit_f32_ceil(F32Ceil), visit_f32_floor(F32Floor), visit_f32_trunc(F32Trunc), visit_f32_nearest(F32Nearest), visit_f32_sqrt(F32Sqrt), visit_f32_add(F32Add), visit_f32_sub(F32Sub), visit_f32_mul(F32Mul), visit_f32_div(F32Div), visit_f32_min(F32Min), visit_f32_max(F32Max), visit_f32_copysign(F32Copysign), visit_f64_abs(F64Abs), visit_f64_neg(F64Neg), visit_f64_ceil(F64Ceil), visit_f64_floor(F64Floor), visit_f64_trunc(F64Trunc), visit_f64_nearest(F64Nearest), visit_f64_sqrt(F64Sqrt), visit_f64_add(F64Add), visit_f64_sub(F64Sub), visit_f64_mul(F64Mul), visit_f64_div(F64Div), visit_f64_min(F64Min), visit_f64_max(F64Max), visit_f64_copysign(F64Copysign), visit_i32_wrap_i64(I32WrapI64), visit_i32_trunc_f32_s(I32TruncF32S), visit_i32_trunc_f32_u(I32TruncF32U), visit_i32_trunc_f64_s(I32TruncF64S), visit_i32_trunc_f64_u(I32TruncF64U), visit_i64_extend_i32_s(I64ExtendI32S), visit_i64_extend_i32_u(I64ExtendI32U), visit_i64_trunc_f32_s(I64TruncF32S), visit_i64_trunc_f32_u(I64TruncF32U), visit_i64_trunc_f64_s(I64TruncF64S), visit_i64_trunc_f64_u(I64TruncF64U), visit_f32_convert_i32_s(F32ConvertI32S), visit_f32_convert_i32_u(F32ConvertI32U), visit_f32_convert_i64_s(F32ConvertI64S), visit_f32_convert_i64_u(F32ConvertI64U), visit_f32_demote_f64(F32DemoteF64), visit_f64_convert_i32_s(F64ConvertI32S), visit_f64_convert_i32_u(F64ConvertI32U), visit_f64_convert_i64_s(F64ConvertI64S), visit_f64_convert_i64_u(F64ConvertI64U), visit_f64_promote_f32(F64PromoteF32), visit_i32_reinterpret_f32(I32ReinterpretF32), visit_i64_reinterpret_f64(I64ReinterpretF64), visit_f32_reinterpret_i32(F32ReinterpretI32), visit_f64_reinterpret_i64(F64ReinterpretI64), + visit_global_get(GlobalGet, u32), visit_i32_const(I32Const, i32), visit_i64_const(I64Const, i64), visit_call(Call, u32), visit_return_call(ReturnCall, u32), visit_memory_size(MemorySize, u32), visit_memory_grow(MemoryGrow, u32), visit_unreachable(Unreachable), visit_nop(Nop), visit_i32_eqz(I32Eqz), visit_i32_eq(I32Eq), visit_i32_ne(I32Ne), visit_i32_lt_s(I32LtS), visit_i32_lt_u(I32LtU), visit_i32_gt_s(I32GtS), visit_i32_gt_u(I32GtU), visit_i32_le_s(I32LeS), visit_i32_le_u(I32LeU), visit_i32_ge_s(I32GeS), visit_i32_ge_u(I32GeU), visit_i64_eqz(I64Eqz), visit_i64_eq(I64Eq), visit_i64_ne(I64Ne), visit_i64_lt_s(I64LtS), visit_i64_lt_u(I64LtU), visit_i64_gt_s(I64GtS), visit_i64_gt_u(I64GtU), visit_i64_le_s(I64LeS), visit_i64_le_u(I64LeU), visit_i64_ge_s(I64GeS), visit_i64_ge_u(I64GeU), visit_f32_eq(F32Eq), visit_f32_ne(F32Ne), visit_f32_lt(F32Lt), visit_f32_gt(F32Gt), visit_f32_le(F32Le), visit_f32_ge(F32Ge), visit_f64_eq(F64Eq), visit_f64_ne(F64Ne), visit_f64_lt(F64Lt), visit_f64_gt(F64Gt), visit_f64_le(F64Le), visit_f64_ge(F64Ge), visit_i32_clz(I32Clz), visit_i32_ctz(I32Ctz), visit_i32_popcnt(I32Popcnt), visit_i32_sub(I32Sub), visit_i32_mul(I32Mul), visit_i32_div_s(I32DivS), visit_i32_div_u(I32DivU), visit_i32_rem_s(I32RemS), visit_i32_rem_u(I32RemU), visit_i32_and(I32And), visit_i32_or(I32Or), visit_i32_xor(I32Xor), visit_i32_shl(I32Shl), visit_i32_shr_s(I32ShrS), visit_i32_shr_u(I32ShrU), visit_i32_rotl(I32Rotl), visit_i32_rotr(I32Rotr), visit_i64_clz(I64Clz), visit_i64_ctz(I64Ctz), visit_i64_popcnt(I64Popcnt), visit_i64_sub(I64Sub), visit_i64_mul(I64Mul), visit_i64_div_s(I64DivS), visit_i64_div_u(I64DivU), visit_i64_rem_s(I64RemS), visit_i64_rem_u(I64RemU), visit_i64_and(I64And), visit_i64_or(I64Or), visit_i64_xor(I64Xor), visit_i64_shl(I64Shl), visit_i64_shr_s(I64ShrS), visit_i64_shr_u(I64ShrU), visit_i64_rotr(I64Rotr), visit_f32_abs(F32Abs), visit_f32_neg(F32Neg), visit_f32_ceil(F32Ceil), visit_f32_floor(F32Floor), visit_f32_trunc(F32Trunc), visit_f32_nearest(F32Nearest), visit_f32_sqrt(F32Sqrt), visit_f32_add(F32Add), visit_f32_sub(F32Sub), visit_f32_mul(F32Mul), visit_f32_div(F32Div), visit_f32_min(F32Min), visit_f32_max(F32Max), visit_f32_copysign(F32Copysign), visit_f64_abs(F64Abs), visit_f64_neg(F64Neg), visit_f64_ceil(F64Ceil), visit_f64_floor(F64Floor), visit_f64_trunc(F64Trunc), visit_f64_nearest(F64Nearest), visit_f64_sqrt(F64Sqrt), visit_f64_add(F64Add), visit_f64_sub(F64Sub), visit_f64_mul(F64Mul), visit_f64_div(F64Div), visit_f64_min(F64Min), visit_f64_max(F64Max), visit_f64_copysign(F64Copysign), visit_i32_wrap_i64(I32WrapI64), visit_i32_trunc_f32_s(I32TruncF32S), visit_i32_trunc_f32_u(I32TruncF32U), visit_i32_trunc_f64_s(I32TruncF64S), visit_i32_trunc_f64_u(I32TruncF64U), visit_i64_extend_i32_s(I64ExtendI32S), visit_i64_extend_i32_u(I64ExtendI32U), visit_i64_trunc_f32_s(I64TruncF32S), visit_i64_trunc_f32_u(I64TruncF32U), visit_i64_trunc_f64_s(I64TruncF64S), visit_i64_trunc_f64_u(I64TruncF64U), visit_f32_convert_i32_s(F32ConvertI32S), visit_f32_convert_i32_u(F32ConvertI32U), visit_f32_convert_i64_s(F32ConvertI64S), visit_f32_convert_i64_u(F32ConvertI64U), visit_f32_demote_f64(F32DemoteF64), visit_f64_convert_i32_s(F64ConvertI32S), visit_f64_convert_i32_u(F64ConvertI32U), visit_f64_convert_i64_s(F64ConvertI64S), visit_f64_convert_i64_u(F64ConvertI64U), visit_f64_promote_f32(F64PromoteF32), visit_i32_reinterpret_f32(I32ReinterpretF32), visit_i64_reinterpret_f64(I64ReinterpretF64), visit_f32_reinterpret_i32(F32ReinterpretI32), visit_f64_reinterpret_i64(F64ReinterpretI64), // sign_extension visit_i32_extend8_s(I32Extend8S), visit_i32_extend16_s(I32Extend16S), visit_i64_extend8_s(I64Extend8S), visit_i64_extend16_s(I64Extend16S), visit_i64_extend32_s(I64Extend32S), @@ -423,6 +419,23 @@ impl<'a, R: WasmModuleResources> wasmparser::VisitOperator<'a> for FunctionBuild } } + fn visit_i32_store(&mut self, memarg: wasmparser::MemArg) -> Self::Output { + let memarg = MemoryArg::new(memarg.offset, memarg.memory); + let len = self.instructions.len(); + if len >= 2 { + let addr = self.instructions[len - 2]; + let value = self.instructions[len - 1]; + if let (Instruction::LocalGet32(addr_local), Instruction::LocalGet32(value_local)) = (addr, value) { + self.instructions.pop(); + self.instructions.pop(); + self.instructions.push(Instruction::I32StoreLocalLocal(memarg, addr_local, value_local)); + return; + } + } + + self.instructions.push(Instruction::I32Store(memarg)); + } + fn visit_drop(&mut self) -> Self::Output { match self.validator.get_operand_type(0) { Some(Some(t)) => self.instructions.push(match t { @@ -447,21 +460,67 @@ impl<'a, R: WasmModuleResources> wasmparser::VisitOperator<'a> for FunctionBuild } fn visit_return(&mut self) -> Self::Output { - if let Some(instr) = self.instructions.last_mut() { - match instr { - Instruction::Call(addr) => { - *instr = Instruction::ReturnCall(*addr); - return; - } - Instruction::CallIndirect(ty, table) => { - *instr = Instruction::ReturnCallIndirect(*ty, *table); - return; - } - _ => {} + self.instructions.push(Instruction::Return); + } + + fn visit_i32_add(&mut self) -> Self::Output { + let len = self.instructions.len(); + if len >= 2 { + let lhs = self.instructions[len - 2]; + let rhs = self.instructions[len - 1]; + if let (Instruction::LocalGet32(a), Instruction::LocalGet32(b)) = (lhs, rhs) { + self.instructions.pop(); + self.instructions.pop(); + self.instructions.push(Instruction::I32AddLocals(a, b)); + return; } } - self.instructions.push(Instruction::Return); + if let Some(Instruction::I32Const(c)) = self.instructions.last().copied() { + self.instructions.pop(); + self.instructions.push(Instruction::I32AddConst(c)); + return; + } + + self.instructions.push(Instruction::I32Add); + } + + fn visit_i64_add(&mut self) -> Self::Output { + let len = self.instructions.len(); + if len >= 2 { + let lhs = self.instructions[len - 2]; + let rhs = self.instructions[len - 1]; + if let (Instruction::LocalGet64(a), Instruction::LocalGet64(b)) = (lhs, rhs) { + self.instructions.pop(); + self.instructions.pop(); + self.instructions.push(Instruction::I64AddLocals(a, b)); + return; + } + } + + if let Some(Instruction::I64Const(c)) = self.instructions.last().copied() { + self.instructions.pop(); + self.instructions.push(Instruction::I64AddConst(c)); + return; + } + + self.instructions.push(Instruction::I64Add); + } + + fn visit_i64_rotl(&mut self) -> Self::Output { + let len = self.instructions.len(); + if len >= 2 { + let lhs = self.instructions[len - 2]; + let rhs = self.instructions[len - 1]; + if let (Instruction::I64Xor, Instruction::I64Const(c)) = (lhs, rhs) { + self.instructions.pop(); + self.instructions.pop(); + self.instructions.push(Instruction::I64XorRotlConst(c)); + return; + } + } + + self.instructions.push(Instruction::I64Rotl); } fn visit_local_get(&mut self, idx: u32) -> Self::Output { @@ -544,6 +603,34 @@ impl<'a, R: WasmModuleResources> wasmparser::VisitOperator<'a> for FunctionBuild return; }; + if let Some(Instruction::I64XorRotlConst(c)) = self.instructions.last().copied() { + match self.validator.get_operand_type(0) { + Some(Some(wasmparser::ValType::I64)) | Some(Some(wasmparser::ValType::F64)) => { + self.instructions.pop(); + self.instructions.push(Instruction::I64XorRotlConstTee(c, resolved_idx)); + return; + } + _ => {} + } + } + + let len = self.instructions.len(); + if len >= 2 { + let addr = self.instructions[len - 2]; + let load = self.instructions[len - 1]; + if let (Instruction::LocalGet32(addr_local), Instruction::I32Load(memarg)) = (addr, load) { + match self.validator.get_operand_type(0) { + Some(Some(wasmparser::ValType::I32)) | Some(Some(wasmparser::ValType::F32)) => { + self.instructions.pop(); + self.instructions.pop(); + self.instructions.push(Instruction::I32LoadLocalTee(memarg, addr_local, resolved_idx)); + return; + } + _ => {} + } + } + } + match self.validator.get_operand_type(0) { Some(Some(t)) => self.instructions.push(match t { wasmparser::ValType::I32 => Instruction::LocalTee32(resolved_idx), @@ -595,6 +682,9 @@ impl<'a, R: WasmModuleResources> wasmparser::VisitOperator<'a> for FunctionBuild ctx.has_else = true; ctx.branch_jumps.push(jump_ip); self.patch_jump_if_zero(cond_jump_ip, self.instructions.len()); + if !matches!(self.instructions.last(), Some(Instruction::Nop)) { + self.instructions.push(Instruction::Nop); + } }; }; } @@ -602,6 +692,9 @@ impl<'a, R: WasmModuleResources> wasmparser::VisitOperator<'a> for FunctionBuild fn visit_end(&mut self) -> Self::Output { if let Some(ctx) = self.ctx_stack.pop() { self.patch_end_jumps(ctx, self.instructions.len()); + if !matches!(self.instructions.last(), Some(Instruction::Nop)) { + self.instructions.push(Instruction::Nop); + } } else { self.instructions.push(Instruction::Return); } diff --git a/crates/tinywasm/benches/argon2id.rs b/crates/tinywasm/benches/argon2id.rs index aa8b38b..ca08378 100644 --- a/crates/tinywasm/benches/argon2id.rs +++ b/crates/tinywasm/benches/argon2id.rs @@ -31,11 +31,11 @@ fn argon2id_run(module: TinyWasmModule) -> Result<()> { fn criterion_benchmark(c: &mut Criterion) { let module = argon2id_parse().expect("argon2id_parse"); - let twasm = argon2id_to_twasm(&module).expect("argon2id_to_twasm"); + let _twasm = argon2id_to_twasm(&module).expect("argon2id_to_twasm"); - c.bench_function("argon2id_parse", |b| b.iter(argon2id_parse)); - c.bench_function("argon2id_to_twasm", |b| b.iter(|| argon2id_to_twasm(&module))); - c.bench_function("argon2id_from_twasm", |b| b.iter(|| argon2id_from_twasm(&twasm))); + // c.bench_function("argon2id_parse", |b| b.iter(argon2id_parse)); + // c.bench_function("argon2id_to_twasm", |b| b.iter(|| argon2id_to_twasm(&module))); + // c.bench_function("argon2id_from_twasm", |b| b.iter(|| argon2id_from_twasm(&twasm))); c.bench_function("argon2id", |b| b.iter(|| argon2id_run(module.clone()))); } diff --git a/crates/tinywasm/benches/tinywasm.rs b/crates/tinywasm/benches/tinywasm.rs index 4848a4a..a949b52 100644 --- a/crates/tinywasm/benches/tinywasm.rs +++ b/crates/tinywasm/benches/tinywasm.rs @@ -33,11 +33,11 @@ fn tinywasm_run(module: TinyWasmModule) -> Result<()> { fn criterion_benchmark(c: &mut Criterion) { let module = tinywasm_parse().expect("tinywasm_parse"); - let twasm = tinywasm_to_twasm(&module).expect("tinywasm_to_twasm"); + let _twasm = tinywasm_to_twasm(&module).expect("tinywasm_to_twasm"); - c.bench_function("tinywasm_parse", |b| b.iter(tinywasm_parse)); - c.bench_function("tinywasm_to_twasm", |b| b.iter(|| tinywasm_to_twasm(&module))); - c.bench_function("tinywasm_from_twasm", |b| b.iter(|| tinywasm_from_twasm(&twasm))); + // c.bench_function("tinywasm_parse", |b| b.iter(tinywasm_parse)); + // c.bench_function("tinywasm_to_twasm", |b| b.iter(|| tinywasm_to_twasm(&module))); + // c.bench_function("tinywasm_from_twasm", |b| b.iter(|| tinywasm_from_twasm(&twasm))); c.bench_function("tinywasm", |b| b.iter(|| tinywasm_run(module.clone()))); } diff --git a/crates/tinywasm/src/interpreter/executor.rs b/crates/tinywasm/src/interpreter/executor.rs index 02e91c5..1da799e 100644 --- a/crates/tinywasm/src/interpreter/executor.rs +++ b/crates/tinywasm/src/interpreter/executor.rs @@ -72,6 +72,10 @@ impl<'store> Executor<'store> { (binary $a:ty, $b:ty => $res:ty, |$lhs:ident, $rhs:ident| $expr:expr) => { self.store.stack.values.binary_diff::<$a, $b, $res>(|$lhs, $rhs| Ok($expr)).to_cf()? }; + (local_set_pop $ty:ty, $local_index:expr) => {{ + let val = self.store.stack.values.pop::<$ty>(); + self.store.stack.values.local_set(&self.cf, *$local_index, val); + }}; } let next = match self.func.instructions.0.get(self.cf.instr_ptr) { @@ -166,58 +170,58 @@ impl<'store> Executor<'store> { return ControlFlow::Continue(()); } Return => return self.exec_return(), - LocalGet32(local_index) => self.store.stack.values.push(self.store.stack.values.local_get_32(&self.cf, *local_index)).to_cf()?, - LocalGet64(local_index) => self.store.stack.values.push(self.store.stack.values.local_get_64(&self.cf, *local_index)).to_cf()?, - LocalGet128(local_index) => self.store.stack.values.push(self.store.stack.values.local_get_128(&self.cf, *local_index)).to_cf()?, - LocalGetRef(local_index) => self.store.stack.values.push(self.store.stack.values.local_get_ref(&self.cf, *local_index)).to_cf()?, - LocalSet32(local_index) => { - let val = self.store.stack.values.pop::(); - self.store.stack.values.local_set_32(&self.cf, *local_index, val); - } - LocalSet64(local_index) => { - let val = self.store.stack.values.pop::(); - self.store.stack.values.local_set_64(&self.cf, *local_index, val); - } - LocalSet128(local_index) => { - let val = self.store.stack.values.pop::(); - self.store.stack.values.local_set_128(&self.cf, *local_index, val); - } - LocalSetRef(local_index) => { - let val = self.store.stack.values.pop::(); - self.store.stack.values.local_set_ref(&self.cf, *local_index, val); - } - LocalCopy32(from, to) => { - let val = self.store.stack.values.local_get_32(&self.cf, *from); - self.store.stack.values.local_set_32(&self.cf, *to, val); - } - LocalCopy64(from, to) => { - let val = self.store.stack.values.local_get_64(&self.cf, *from); - self.store.stack.values.local_set_64(&self.cf, *to, val); - } - LocalCopy128(from, to) => { - let val = self.store.stack.values.local_get_128(&self.cf, *from); - self.store.stack.values.local_set_128(&self.cf, *to, val); - } - LocalCopyRef(from, to) => { - let val = self.store.stack.values.local_get_ref(&self.cf, *from); - self.store.stack.values.local_set_ref(&self.cf, *to, val); - } - LocalTee32(local_index) => { - let val = self.store.stack.values.peek::(); - self.store.stack.values.local_set_32(&self.cf, *local_index, val); - } - LocalTee64(local_index) => { - let val = self.store.stack.values.peek::(); - self.store.stack.values.local_set_64(&self.cf, *local_index, val); + LocalGet32(local_index) => self.store.stack.values.push(self.store.stack.values.local_get::(&self.cf, *local_index)).to_cf()?, + LocalGet64(local_index) => self.store.stack.values.push(self.store.stack.values.local_get::(&self.cf, *local_index)).to_cf()?, + LocalGet128(local_index) => self.store.stack.values.push(self.store.stack.values.local_get::(&self.cf, *local_index)).to_cf()?, + LocalGetRef(local_index) => self.store.stack.values.push(self.store.stack.values.local_get::(&self.cf, *local_index)).to_cf()?, + LocalSet32(local_index) => stack_op!(local_set_pop Value32, local_index), + LocalSet64(local_index) => stack_op!(local_set_pop Value64, local_index), + LocalSet128(local_index) => stack_op!(local_set_pop Value128, local_index), + LocalSetRef(local_index) => stack_op!(local_set_pop ValueRef, local_index), + LocalCopy32(from, to) => self.store.stack.values.local_set(&self.cf, *to, self.store.stack.values.local_get::(&self.cf, *from)), + LocalCopy64(from, to) => self.store.stack.values.local_set(&self.cf, *to, self.store.stack.values.local_get::(&self.cf, *from)), + LocalCopy128(from, to) => self.store.stack.values.local_set(&self.cf, *to, self.store.stack.values.local_get::(&self.cf, *from)), + LocalCopyRef(from, to) => self.store.stack.values.local_set(&self.cf, *to, self.store.stack.values.local_get::(&self.cf, *from)), + I32AddLocals(a, b) => self.store.stack.values.push( + self.store.stack.values.local_get::(&self.cf, *a).wrapping_add(self.store.stack.values.local_get::(&self.cf, *b)), + ).to_cf()?, + I64AddLocals(a, b) => self.store.stack.values.push( + self.store.stack.values.local_get::(&self.cf, *a).wrapping_add(self.store.stack.values.local_get::(&self.cf, *b)), + ).to_cf()?, + I32AddConst(c) => stack_op!(unary i32, |v| v.wrapping_add(*c)), + I64AddConst(c) => stack_op!(unary i64, |v| v.wrapping_add(*c)), + I32StoreLocalLocal(m, addr_local, value_local) => { + let mem = self.store.state.get_mem_mut(self.module.resolve_mem_addr(m.mem_addr())); + let addr = u64::from(self.store.stack.values.local_get::(&self.cf, *addr_local)); + let value = self.store.stack.values.local_get::(&self.cf, *value_local).to_mem_bytes(); + if let Err(e) = mem.store((m.offset() + addr) as usize, value.len(), &value) { + return ControlFlow::Break(Some(e)); + } } - LocalTee128(local_index) => { - let val = self.store.stack.values.peek::(); - self.store.stack.values.local_set_128(&self.cf, *local_index, val); + I32LoadLocalTee(m, addr_local, dst_local) => { + let mem = self.store.state.get_mem(self.module.resolve_mem_addr(m.mem_addr())); + let addr = u64::from(self.store.stack.values.local_get::(&self.cf, *addr_local)); + let Some(Ok(addr)) = m.offset().checked_add(addr).map(|a| a.try_into()) else { + return ControlFlow::Break(Some(Error::Trap(Trap::MemoryOutOfBounds { + offset: addr as usize, + len: 4, + max: 0, + }))); + }; + let value = mem.load_as::<4, i32>(addr).to_cf()?; + self.store.stack.values.local_set(&self.cf, *dst_local, value); + self.store.stack.values.push(value).to_cf()?; } - LocalTeeRef(local_index) => { - let val = self.store.stack.values.peek::(); - self.store.stack.values.local_set_ref(&self.cf, *local_index, val); + I64XorRotlConst(c) => stack_op!(binary i64, |lhs, rhs| (lhs ^ rhs).rotate_left(*c as u32)), + I64XorRotlConstTee(c, local_index) => { + self.store.stack.values.binary_same::(|lhs, rhs| Ok((lhs ^ rhs).rotate_left(*c as u32))).to_cf()?; + let val = self.store.stack.values.peek::(); + self.store.stack.values.local_set(&self.cf, *local_index, val); } + LocalTee32(local_index) => self.store.stack.values.local_set(&self.cf, *local_index, self.store.stack.values.peek::()), + LocalTee64(local_index) => self.store.stack.values.local_set(&self.cf, *local_index, self.store.stack.values.peek::()), + LocalTee128(local_index) => self.store.stack.values.local_set(&self.cf, *local_index, self.store.stack.values.peek::()), + LocalTeeRef(local_index) => self.store.stack.values.local_set(&self.cf, *local_index, self.store.stack.values.peek::()), GlobalGet(global_index) => self.exec_global_get(*global_index).to_cf()?, GlobalSet32(global_index) => self.exec_global_set::(*global_index), GlobalSet64(global_index) => self.exec_global_set::(*global_index), diff --git a/crates/tinywasm/src/interpreter/stack/value_stack.rs b/crates/tinywasm/src/interpreter/stack/value_stack.rs index 76c4b28..b461f7a 100644 --- a/crates/tinywasm/src/interpreter/stack/value_stack.rs +++ b/crates/tinywasm/src/interpreter/stack/value_stack.rs @@ -268,43 +268,13 @@ impl ValueStack { } #[inline] - pub(crate) fn local_get_32(&self, frame: &CallFrame, index: LocalAddr) -> Value32 { - self.stack_32.get(frame.locals_base.s32 + index as usize) + pub(crate) fn local_get(&self, frame: &CallFrame, index: LocalAddr) -> T { + T::local_get(self, frame, index) } #[inline] - pub(crate) fn local_get_64(&self, frame: &CallFrame, index: LocalAddr) -> Value64 { - self.stack_64.get(frame.locals_base.s64 + index as usize) - } - - #[inline] - pub(crate) fn local_get_128(&self, frame: &CallFrame, index: LocalAddr) -> Value128 { - self.stack_128.get(frame.locals_base.s128 + index as usize) - } - - #[inline] - pub(crate) fn local_get_ref(&self, frame: &CallFrame, index: LocalAddr) -> ValueRef { - self.stack_ref.get(frame.locals_base.sref + index as usize) - } - - #[inline] - pub(crate) fn local_set_32(&mut self, frame: &CallFrame, index: LocalAddr, value: Value32) { - self.stack_32.set(frame.locals_base.s32 + index as usize, value); - } - - #[inline] - pub(crate) fn local_set_64(&mut self, frame: &CallFrame, index: LocalAddr, value: Value64) { - self.stack_64.set(frame.locals_base.s64 + index as usize, value); - } - - #[inline] - pub(crate) fn local_set_128(&mut self, frame: &CallFrame, index: LocalAddr, value: Value128) { - self.stack_128.set(frame.locals_base.s128 + index as usize, value); - } - - #[inline] - pub(crate) fn local_set_ref(&mut self, frame: &CallFrame, index: LocalAddr, value: ValueRef) { - self.stack_ref.set(frame.locals_base.sref + index as usize, value); + pub(crate) fn local_set(&mut self, frame: &CallFrame, index: LocalAddr, value: T) { + T::local_set(self, frame, index, value); } pub(crate) fn push_dyn(&mut self, value: TinyWasmValue) -> Result<()> { diff --git a/crates/tinywasm/src/interpreter/values.rs b/crates/tinywasm/src/interpreter/values.rs index b1530e7..9b6e2c1 100644 --- a/crates/tinywasm/src/interpreter/values.rs +++ b/crates/tinywasm/src/interpreter/values.rs @@ -1,6 +1,7 @@ use crate::{Result, interpreter::value128::Value128}; -use super::stack::ValueStack; +use super::stack::{CallFrame, ValueStack}; +use tinywasm_types::LocalAddr; use tinywasm_types::{ExternRef, FuncRef, ValType, WasmValue}; pub(crate) type Value32 = u32; @@ -105,6 +106,12 @@ mod sealed { pub(crate) trait InternalValue: sealed::Sealed + Into { fn stack_push(stack: &mut ValueStack, value: Self) -> Result<()>; + fn local_get(stack: &ValueStack, frame: &CallFrame, index: LocalAddr) -> Self + where + Self: Sized; + fn local_set(stack: &mut ValueStack, frame: &CallFrame, index: LocalAddr, value: Self) + where + Self: Sized; fn replace_top(stack: &mut ValueStack, func: impl FnOnce(Self) -> Result) -> Result<()> where Self: Sized; @@ -124,7 +131,7 @@ pub(crate) trait InternalValue: sealed::Sealed + Into { } macro_rules! impl_internalvalue { - ($( $variant:ident, $stack:ident, $internal:ty, $outer:ty, $to_internal:expr, $to_outer:expr )*) => { + ($( $variant:ident, $stack:ident, $stack_base:ident, $outer:ty, $to_internal:expr, $to_outer:expr )*) => { $( impl sealed::Sealed for $outer {} @@ -140,6 +147,16 @@ macro_rules! impl_internalvalue { stack.$stack.push($to_internal(value)) } + #[inline] + fn local_get(stack: &ValueStack, frame: &CallFrame, index: LocalAddr) -> Self { + $to_outer(stack.$stack.get(frame.locals_base.$stack_base + index as usize)) + } + + #[inline] + fn local_set(stack: &mut ValueStack, frame: &CallFrame, index: LocalAddr, value: Self) { + stack.$stack.set(frame.locals_base.$stack_base + index as usize, $to_internal(value)); + } + #[inline] fn stack_pop(stack: &mut ValueStack) -> Self { $to_outer(stack.$stack.pop()) @@ -179,12 +196,12 @@ macro_rules! impl_internalvalue { } impl_internalvalue! { - Value32, stack_32, u32, u32, |v| v, |v| v - Value64, stack_64, u64, u64, |v| v, |v| v - Value32, stack_32, u32, i32, |v: i32| u32::from_ne_bytes(v.to_ne_bytes()), |v: u32| i32::from_ne_bytes(v.to_ne_bytes()) - Value64, stack_64, u64, i64, |v: i64| u64::from_ne_bytes(v.to_ne_bytes()), |v: u64| i64::from_ne_bytes(v.to_ne_bytes()) - Value32, stack_32, u32, f32, f32::to_bits, f32::from_bits - Value64, stack_64, u64, f64, f64::to_bits, f64::from_bits - ValueRef, stack_ref, ValueRef, ValueRef, |v| v, |v| v - Value128, stack_128, Value128, Value128, |v| v, |v| v + Value32, stack_32, s32, u32, |v| v, |v| v + Value64, stack_64, s64, u64, |v| v, |v| v + Value32, stack_32, s32, i32, |v: i32| u32::from_ne_bytes(v.to_ne_bytes()), |v: u32| i32::from_ne_bytes(v.to_ne_bytes()) + Value64, stack_64, s64, i64, |v: i64| u64::from_ne_bytes(v.to_ne_bytes()), |v: u64| i64::from_ne_bytes(v.to_ne_bytes()) + Value32, stack_32, s32, f32, f32::to_bits, f32::from_bits + Value64, stack_64, s64, f64, f64::to_bits, f64::from_bits + ValueRef, stack_ref, sref, ValueRef, |v| v, |v| v + Value128, stack_128, s128, Value128, |v| v, |v| v } diff --git a/crates/types/src/instructions.rs b/crates/types/src/instructions.rs index 71b518f..7f0cef5 100644 --- a/crates/types/src/instructions.rs +++ b/crates/types/src/instructions.rs @@ -51,7 +51,12 @@ pub enum ConstInstruction { #[rustfmt::skip] pub enum Instruction { LocalCopy32(LocalAddr, LocalAddr), LocalCopy64(LocalAddr, LocalAddr), LocalCopy128(LocalAddr, LocalAddr), LocalCopyRef(LocalAddr, LocalAddr), - + I32AddLocals(LocalAddr, LocalAddr), I64AddLocals(LocalAddr, LocalAddr), + I32AddConst(i32), I64AddConst(i64), + I32StoreLocalLocal(MemoryArg, LocalAddr, LocalAddr), + I32LoadLocalTee(MemoryArg, LocalAddr, LocalAddr), + I64XorRotlConst(i64), + I64XorRotlConstTee(i64, LocalAddr), // > Control Instructions (jump-oriented, lowered from structured control during parsing) // See Unreachable, -- cgit v1.3.1