cranelift: Add heap support to the interpreter (#3302)

* cranelift: Add heaps to interpreter

* cranelift: Add RunTest Environment mechanism to  test interpret

* cranelift: Remove unused `MemoryError`

* cranelift: Add docs for `State::resolve_global_value`

* cranelift: Rename heap tests

* cranelift: Refactor heap address resolution

* Fix typos and clarify docs (thanks @cfallin)
This commit is contained in:
Afonso Bordado
2022-07-05 17:05:26 +01:00
committed by GitHub
parent 76a2545a7f
commit e91f493ff5
7 changed files with 531 additions and 140 deletions

View File

@@ -1,10 +1,11 @@
test interpret
test run test run
target x86_64 target x86_64
target s390x target s390x
target aarch64 target aarch64
function %static_heap_i64_load_store(i64 vmctx, i64, i32) -> i32 { function %static_heap_i64(i64 vmctx, i64, i32) -> i32 {
gv0 = vmctx gv0 = vmctx
gv1 = load.i64 notrap aligned gv0+0 gv1 = load.i64 notrap aligned gv0+0
heap0 = static gv1, min 0x1000, bound 0x1_0000_0000, offset_guard 0, index_type i64 heap0 = static gv1, min 0x1000, bound 0x1_0000_0000, offset_guard 0, index_type i64
@@ -16,13 +17,13 @@ block0(v0: i64, v1: i64, v2: i32):
return v4 return v4
} }
; heap: static, size=0x1000, ptr=vmctx+0, bound=vmctx+8 ; heap: static, size=0x1000, ptr=vmctx+0, bound=vmctx+8
; run: %static_heap_i64_load_store(0, 1) == 1 ; run: %static_heap_i64(0, 1) == 1
; run: %static_heap_i64_load_store(0, -1) == -1 ; run: %static_heap_i64(0, -1) == -1
; run: %static_heap_i64_load_store(16, 1) == 1 ; run: %static_heap_i64(16, 1) == 1
; run: %static_heap_i64_load_store(16, -1) == -1 ; run: %static_heap_i64(16, -1) == -1
function %static_heap_i32_load_store(i64 vmctx, i32, i32) -> i32 { function %static_heap_i32(i64 vmctx, i32, i32) -> i32 {
gv0 = vmctx gv0 = vmctx
gv1 = load.i64 notrap aligned gv0+0 gv1 = load.i64 notrap aligned gv0+0
heap0 = static gv1, min 0x1000, bound 0x1_0000_0000, offset_guard 0, index_type i32 heap0 = static gv1, min 0x1000, bound 0x1_0000_0000, offset_guard 0, index_type i32
@@ -34,13 +35,13 @@ block0(v0: i64, v1: i32, v2: i32):
return v4 return v4
} }
; heap: static, size=0x1000, ptr=vmctx+0, bound=vmctx+8 ; heap: static, size=0x1000, ptr=vmctx+0, bound=vmctx+8
; run: %static_heap_i32_load_store(0, 1) == 1 ; run: %static_heap_i32(0, 1) == 1
; run: %static_heap_i32_load_store(0, -1) == -1 ; run: %static_heap_i32(0, -1) == -1
; run: %static_heap_i32_load_store(16, 1) == 1 ; run: %static_heap_i32(16, 1) == 1
; run: %static_heap_i32_load_store(16, -1) == -1 ; run: %static_heap_i32(16, -1) == -1
function %static_heap_i32_load_store_no_min(i64 vmctx, i32, i32) -> i32 { function %heap_no_min(i64 vmctx, i32, i32) -> i32 {
gv0 = vmctx gv0 = vmctx
gv1 = load.i64 notrap aligned gv0+0 gv1 = load.i64 notrap aligned gv0+0
heap0 = static gv1, bound 0x1_0000_0000, offset_guard 0, index_type i32 heap0 = static gv1, bound 0x1_0000_0000, offset_guard 0, index_type i32
@@ -52,13 +53,13 @@ block0(v0: i64, v1: i32, v2: i32):
return v4 return v4
} }
; heap: static, size=0x1000, ptr=vmctx+0, bound=vmctx+8 ; heap: static, size=0x1000, ptr=vmctx+0, bound=vmctx+8
; run: %static_heap_i32_load_store_no_min(0, 1) == 1 ; run: %heap_no_min(0, 1) == 1
; run: %static_heap_i32_load_store_no_min(0, -1) == -1 ; run: %heap_no_min(0, -1) == -1
; run: %static_heap_i32_load_store_no_min(16, 1) == 1 ; run: %heap_no_min(16, 1) == 1
; run: %static_heap_i32_load_store_no_min(16, -1) == -1 ; run: %heap_no_min(16, -1) == -1
function %dynamic_heap_i64_load_store(i64 vmctx, i64, i32) -> i32 { function %dynamic_i64(i64 vmctx, i64, i32) -> i32 {
gv0 = vmctx gv0 = vmctx
gv1 = load.i64 notrap aligned gv0+0 gv1 = load.i64 notrap aligned gv0+0
gv2 = load.i64 notrap aligned gv0+8 gv2 = load.i64 notrap aligned gv0+8
@@ -71,13 +72,13 @@ block0(v0: i64, v1: i64, v2: i32):
return v4 return v4
} }
; heap: dynamic, size=0x1000, ptr=vmctx+0, bound=vmctx+8 ; heap: dynamic, size=0x1000, ptr=vmctx+0, bound=vmctx+8
; run: %dynamic_heap_i64_load_store(0, 1) == 1 ; run: %dynamic_i64(0, 1) == 1
; run: %dynamic_heap_i64_load_store(0, -1) == -1 ; run: %dynamic_i64(0, -1) == -1
; run: %dynamic_heap_i64_load_store(16, 1) == 1 ; run: %dynamic_i64(16, 1) == 1
; run: %dynamic_heap_i64_load_store(16, -1) == -1 ; run: %dynamic_i64(16, -1) == -1
function %dynamic_heap_i32_load_store(i64 vmctx, i32, i32) -> i32 { function %dynamic_i32(i64 vmctx, i32, i32) -> i32 {
gv0 = vmctx gv0 = vmctx
gv1 = load.i64 notrap aligned gv0+0 gv1 = load.i64 notrap aligned gv0+0
gv2 = load.i64 notrap aligned gv0+8 gv2 = load.i64 notrap aligned gv0+8
@@ -90,13 +91,13 @@ block0(v0: i64, v1: i32, v2: i32):
return v4 return v4
} }
; heap: dynamic, size=0x1000, ptr=vmctx+0, bound=vmctx+8 ; heap: dynamic, size=0x1000, ptr=vmctx+0, bound=vmctx+8
; run: %dynamic_heap_i32_load_store(0, 1) == 1 ; run: %dynamic_i32(0, 1) == 1
; run: %dynamic_heap_i32_load_store(0, -1) == -1 ; run: %dynamic_i32(0, -1) == -1
; run: %dynamic_heap_i32_load_store(16, 1) == 1 ; run: %dynamic_i32(16, 1) == 1
; run: %dynamic_heap_i32_load_store(16, -1) == -1 ; run: %dynamic_i32(16, -1) == -1
function %multi_heap_load_store(i64 vmctx, i32, i32) -> i32 { function %multi_load_store(i64 vmctx, i32, i32) -> i32 {
gv0 = vmctx gv0 = vmctx
gv1 = load.i64 notrap aligned gv0+0 gv1 = load.i64 notrap aligned gv0+0
gv2 = load.i64 notrap aligned gv0+16 gv2 = load.i64 notrap aligned gv0+16
@@ -125,12 +126,47 @@ block0(v0: i64, v1: i32, v2: i32):
} }
; heap: static, size=0x1000, ptr=vmctx+0, bound=vmctx+8 ; heap: static, size=0x1000, ptr=vmctx+0, bound=vmctx+8
; heap: dynamic, size=0x1000, ptr=vmctx+16, bound=vmctx+24 ; heap: dynamic, size=0x1000, ptr=vmctx+16, bound=vmctx+24
; run: %multi_heap_load_store(1, 2) == 3 ; run: %multi_load_store(1, 2) == 3
; run: %multi_heap_load_store(4, 5) == 9 ; run: %multi_load_store(4, 5) == 9
function %static_heap_i64_load_store_unaligned(i64 vmctx, i64, i32) -> i32 { ; Uses multiple heaps, but heap0 refers to the second heap, and heap1 refers to the first heap
; This is a regression test for the interpreter
function %out_of_order(i64 vmctx, i32, i32) -> i32 {
gv0 = vmctx
gv1 = load.i64 notrap aligned gv0+0
gv2 = load.i64 notrap aligned gv0+16
gv3 = load.i64 notrap aligned gv0+24
heap0 = dynamic gv2, bound gv3, offset_guard 0, index_type i32
heap1 = static gv1, min 0x1000, bound 0x1_0000_0000, offset_guard 0, index_type i64
block0(v0: i64, v1: i32, v2: i32):
v3 = iconst.i32 0
v4 = iconst.i64 0
; Store lhs in heap0
v5 = heap_addr.i64 heap0, v3, 4
store.i32 v1, v5
; Store rhs in heap1
v6 = heap_addr.i64 heap1, v4, 4
store.i32 v2, v6
v7 = load.i32 v5
v8 = load.i32 v6
v9 = iadd.i32 v7, v8
return v9
}
; heap: static, size=0x1000, ptr=vmctx+0, bound=vmctx+8
; heap: dynamic, size=0x1000, ptr=vmctx+16, bound=vmctx+24
; run: %out_of_order(1, 2) == 3
; run: %out_of_order(4, 5) == 9
function %unaligned_access(i64 vmctx, i64, i32) -> i32 {
gv0 = vmctx gv0 = vmctx
gv1 = load.i64 notrap aligned gv0+0 gv1 = load.i64 notrap aligned gv0+0
heap0 = static gv1, min 0x1000, bound 0x1_0000_0000, offset_guard 0, index_type i64 heap0 = static gv1, min 0x1000, bound 0x1_0000_0000, offset_guard 0, index_type i64
@@ -142,18 +178,18 @@ block0(v0: i64, v1: i64, v2: i32):
return v4 return v4
} }
; heap: static, size=0x1000, ptr=vmctx+0, bound=vmctx+8 ; heap: static, size=0x1000, ptr=vmctx+0, bound=vmctx+8
; run: %static_heap_i64_load_store_unaligned(0, 1) == 1 ; run: %unaligned_access(0, 1) == 1
; run: %static_heap_i64_load_store_unaligned(0, -1) == -1 ; run: %unaligned_access(0, -1) == -1
; run: %static_heap_i64_load_store_unaligned(1, 1) == 1 ; run: %unaligned_access(1, 1) == 1
; run: %static_heap_i64_load_store_unaligned(1, -1) == -1 ; run: %unaligned_access(1, -1) == -1
; run: %static_heap_i64_load_store_unaligned(2, 1) == 1 ; run: %unaligned_access(2, 1) == 1
; run: %static_heap_i64_load_store_unaligned(2, -1) == -1 ; run: %unaligned_access(2, -1) == -1
; run: %static_heap_i64_load_store_unaligned(3, 1) == 1 ; run: %unaligned_access(3, 1) == 1
; run: %static_heap_i64_load_store_unaligned(3, -1) == -1 ; run: %unaligned_access(3, -1) == -1
; This stores data in the place of the pointer in the vmctx struct, not in the heap itself. ; This stores data in the place of the pointer in the vmctx struct, not in the heap itself.
function %static_heap_i64_iadd_imm(i64 vmctx, i32) -> i32 { function %iadd_imm(i64 vmctx, i32) -> i32 {
gv0 = vmctx gv0 = vmctx
gv1 = iadd_imm.i64 gv0, 0 gv1 = iadd_imm.i64 gv0, 0
heap0 = static gv1, min 0x1000, bound 0x1_0000_0000, offset_guard 0x8000_0000, index_type i64 heap0 = static gv1, min 0x1000, bound 0x1_0000_0000, offset_guard 0x8000_0000, index_type i64
@@ -166,5 +202,5 @@ block0(v0: i64, v1: i32):
return v4 return v4
} }
; heap: static, size=0x1000, ptr=vmctx+0, bound=vmctx+8 ; heap: static, size=0x1000, ptr=vmctx+0, bound=vmctx+8
; run: %static_heap_i64_iadd_imm(1) == 1 ; run: %iadd_imm(1) == 1
; run: %static_heap_i64_iadd_imm(-1) == -1 ; run: %iadd_imm(-1) == -1

View File

@@ -1,6 +1,5 @@
use anyhow::anyhow; use anyhow::anyhow;
use cranelift_codegen::data_value::DataValue; use cranelift_codegen::ir::{ArgumentPurpose, Function};
use cranelift_codegen::ir::Type;
use cranelift_reader::parse_heap_command; use cranelift_reader::parse_heap_command;
use cranelift_reader::{Comment, HeapCommand}; use cranelift_reader::{Comment, HeapCommand};
@@ -45,68 +44,37 @@ impl RuntestEnvironment {
!self.heaps.is_empty() !self.heaps.is_empty()
} }
/// Allocates a struct to be injected into the test. /// Allocates memory for heaps
pub fn runtime_struct(&self) -> RuntestContext { pub fn allocate_memory(&self) -> Vec<HeapMemory> {
RuntestContext::new(&self) self.heaps
}
}
type HeapMemory = Vec<u8>;
/// A struct that provides info about the environment to the test
#[derive(Debug, Clone)]
pub struct RuntestContext {
/// Store the heap memory alongside the context info so that we don't accidentally deallocate
/// it too early.
#[allow(dead_code)]
heaps: Vec<HeapMemory>,
/// This is the actual struct that gets passed into the `vmctx` argument of the tests.
/// It has a specific memory layout that all tests agree with.
///
/// Currently we only have to store heap info, so we store the heap start and end addresses in
/// a 64 bit slot for each heap.
///
/// ┌────────────┐
/// │heap0: start│
/// ├────────────┤
/// │heap0: end │
/// ├────────────┤
/// │heap1: start│
/// ├────────────┤
/// │heap1: end │
/// ├────────────┤
/// │etc... │
/// └────────────┘
context_struct: Vec<u64>,
}
impl RuntestContext {
pub fn new(env: &RuntestEnvironment) -> Self {
let heaps: Vec<HeapMemory> = env
.heaps
.iter() .iter()
.map(|cmd| { .map(|cmd| {
let size: u64 = cmd.size.into(); let size: u64 = cmd.size.into();
vec![0u8; size as usize] vec![0u8; size as usize]
}) })
.collect(); .collect()
let context_struct = heaps
.iter()
.flat_map(|heap| [heap.as_ptr(), heap.as_ptr().wrapping_add(heap.len())])
.map(|p| p as usize as u64)
.collect();
Self {
heaps,
context_struct,
}
} }
/// Creates a [DataValue] with a target isa pointer type to the context struct. /// Validates the signature of a [Function] ensuring that if this environment is active, the
pub fn pointer(&self, ty: Type) -> DataValue { /// function has a `vmctx` argument
let ptr = self.context_struct.as_ptr() as usize as i128; pub fn validate_signature(&self, func: &Function) -> Result<(), String> {
DataValue::from_integer(ptr, ty).expect("Failed to cast pointer to native target size") let first_arg_is_vmctx = func
.signature
.params
.first()
.map(|p| p.purpose == ArgumentPurpose::VMContext)
.unwrap_or(false);
if !first_arg_is_vmctx && self.is_active() {
return Err(concat!(
"This test requests a heap, but the first argument is not `i64 vmctx`.\n",
"See docs/testing.md for more info on using heap annotations."
)
.to_string());
}
Ok(())
} }
} }
pub(crate) type HeapMemory = Vec<u8>;

View File

@@ -3,10 +3,13 @@
//! The `interpret` test command interprets each function on the host machine //! The `interpret` test command interprets each function on the host machine
//! using [RunCommand](cranelift_reader::RunCommand)s. //! using [RunCommand](cranelift_reader::RunCommand)s.
use crate::runtest_environment::RuntestEnvironment;
use crate::subtest::{Context, SubTest}; use crate::subtest::{Context, SubTest};
use cranelift_codegen::data_value::DataValue;
use cranelift_codegen::ir::types::I64;
use cranelift_codegen::{self, ir}; use cranelift_codegen::{self, ir};
use cranelift_interpreter::environment::FunctionStore; use cranelift_interpreter::environment::FunctionStore;
use cranelift_interpreter::interpreter::{Interpreter, InterpreterState}; use cranelift_interpreter::interpreter::{HeapInit, Interpreter, InterpreterState};
use cranelift_interpreter::step::ControlFlow; use cranelift_interpreter::step::ControlFlow;
use cranelift_reader::{parse_run_command, TestCommand}; use cranelift_reader::{parse_run_command, TestCommand};
use log::trace; use log::trace;
@@ -36,6 +39,7 @@ impl SubTest for TestInterpret {
} }
fn run(&self, func: Cow<ir::Function>, context: &Context) -> anyhow::Result<()> { fn run(&self, func: Cow<ir::Function>, context: &Context) -> anyhow::Result<()> {
let test_env = RuntestEnvironment::parse(&context.details.comments[..])?;
for comment in context.details.comments.iter() { for comment in context.details.comments.iter() {
if let Some(command) = parse_run_command(comment.text, &func.signature)? { if let Some(command) = parse_run_command(comment.text, &func.signature)? {
trace!("Parsed run command: {}", command); trace!("Parsed run command: {}", command);
@@ -44,11 +48,21 @@ impl SubTest for TestInterpret {
env.add(func.name.to_string(), &func); env.add(func.name.to_string(), &func);
command command
.run(|func_name, args| { .run(|func_name, run_args| {
test_env.validate_signature(&func)?;
let mut state = InterpreterState::default().with_function_store(env);
let mut args = Vec::with_capacity(run_args.len());
if test_env.is_active() {
let vmctx_addr = register_heaps(&mut state, &test_env);
args.push(vmctx_addr);
}
args.extend_from_slice(run_args);
// Because we have stored function names with a leading %, we need to re-add it. // Because we have stored function names with a leading %, we need to re-add it.
let func_name = &format!("%{}", func_name); let func_name = &format!("%{}", func_name);
let state = InterpreterState::default().with_function_store(env); match Interpreter::new(state).call_by_name(func_name, &args) {
match Interpreter::new(state).call_by_name(func_name, args) {
Ok(ControlFlow::Return(results)) => Ok(results.to_vec()), Ok(ControlFlow::Return(results)) => Ok(results.to_vec()),
Ok(_) => { Ok(_) => {
panic!("Unexpected returned control flow--this is likely a bug.") panic!("Unexpected returned control flow--this is likely a bug.")
@@ -62,3 +76,34 @@ impl SubTest for TestInterpret {
Ok(()) Ok(())
} }
} }
/// Build a VMContext struct with the layout described in docs/testing.md.
pub fn register_heaps<'a>(
state: &mut InterpreterState<'a>,
test_env: &RuntestEnvironment,
) -> DataValue {
let mem = test_env.allocate_memory();
let vmctx_struct = mem
.into_iter()
// This memory layout (a contiguous list of base + bound ptrs)
// is enforced by the RuntestEnvironment when parsing the heap
// directives. So we are safe to replicate that here.
.flat_map(|mem| {
let heap_len = mem.len() as u64;
let heap = state.register_heap(HeapInit::FromBacking(mem));
[
state.get_heap_address(I64, heap, 0).unwrap(),
state.get_heap_address(I64, heap, heap_len).unwrap(),
]
})
.map(|addr| {
let mut mem = [0u8; 8];
addr.write_to_slice(&mut mem[..]);
mem
})
.flatten()
.collect();
let vmctx_heap = state.register_heap(HeapInit::FromBacking(vmctx_struct));
state.get_heap_address(I64, vmctx_heap, 0).unwrap()
}

View File

@@ -3,10 +3,11 @@
//! The `run` test command compiles each function on the host machine and executes it //! The `run` test command compiles each function on the host machine and executes it
use crate::function_runner::SingleFunctionCompiler; use crate::function_runner::SingleFunctionCompiler;
use crate::runtest_environment::RuntestEnvironment; use crate::runtest_environment::{HeapMemory, RuntestEnvironment};
use crate::subtest::{Context, SubTest}; use crate::subtest::{Context, SubTest};
use cranelift_codegen::data_value::DataValue;
use cranelift_codegen::ir; use cranelift_codegen::ir;
use cranelift_codegen::ir::ArgumentPurpose; use cranelift_codegen::ir::Type;
use cranelift_reader::parse_run_command; use cranelift_reader::parse_run_command;
use cranelift_reader::TestCommand; use cranelift_reader::TestCommand;
use log::trace; use log::trace;
@@ -64,25 +65,13 @@ impl SubTest for TestRun {
let compiled_fn = compiler.compile(func.clone().into_owned())?; let compiled_fn = compiler.compile(func.clone().into_owned())?;
command command
.run(|_, run_args| { .run(|_, run_args| {
let runtime_struct = test_env.runtime_struct(); test_env.validate_signature(&func)?;
let (_heaps, _ctx_struct, vmctx_ptr) =
let first_arg_is_vmctx = func build_vmctx_struct(&test_env, context.isa.unwrap().pointer_type());
.signature
.params
.first()
.map(|p| p.purpose == ArgumentPurpose::VMContext)
.unwrap_or(false);
if !first_arg_is_vmctx && test_env.is_active() {
return Err(concat!(
"This test requests a heap, but the first argument is not `i64 vmctx`.\n",
"See docs/testing.md for more info on using heap annotations."
).to_string());
}
let mut args = Vec::with_capacity(run_args.len()); let mut args = Vec::with_capacity(run_args.len());
if test_env.is_active() { if test_env.is_active() {
args.push(runtime_struct.pointer(context.isa.unwrap().pointer_type())); args.push(vmctx_ptr);
} }
args.extend_from_slice(run_args); args.extend_from_slice(run_args);
@@ -94,3 +83,24 @@ impl SubTest for TestRun {
Ok(()) Ok(())
} }
} }
/// Build a VMContext struct with the layout described in docs/testing.md.
pub fn build_vmctx_struct(
test_env: &RuntestEnvironment,
ptr_ty: Type,
) -> (Vec<HeapMemory>, Vec<u64>, DataValue) {
let heaps = test_env.allocate_memory();
let context_struct: Vec<u64> = heaps
.iter()
.flat_map(|heap| [heap.as_ptr(), heap.as_ptr().wrapping_add(heap.len())])
.map(|p| p as usize as u64)
.collect();
let ptr = context_struct.as_ptr() as usize as i128;
let ptr_dv =
DataValue::from_integer(ptr, ptr_ty).expect("Failed to cast pointer to native target size");
// Return all these to make sure we don't deallocate the heaps too early
(heaps, context_struct, ptr_dv)
}

View File

@@ -8,12 +8,16 @@ use crate::frame::Frame;
use crate::instruction::DfgInstructionContext; use crate::instruction::DfgInstructionContext;
use crate::state::{MemoryError, State}; use crate::state::{MemoryError, State};
use crate::step::{step, ControlFlow, StepError}; use crate::step::{step, ControlFlow, StepError};
use crate::value::ValueError; use crate::value::{Value, ValueError};
use cranelift_codegen::data_value::DataValue; use cranelift_codegen::data_value::DataValue;
use cranelift_codegen::ir::condcodes::{FloatCC, IntCC}; use cranelift_codegen::ir::condcodes::{FloatCC, IntCC};
use cranelift_codegen::ir::{Block, FuncRef, Function, StackSlot, Type, Value as ValueRef}; use cranelift_codegen::ir::{
ArgumentPurpose, Block, FuncRef, Function, GlobalValue, GlobalValueData, Heap, StackSlot, Type,
Value as ValueRef,
};
use log::trace; use log::trace;
use std::collections::HashSet; use std::collections::HashSet;
use std::convert::{TryFrom, TryInto};
use std::fmt::Debug; use std::fmt::Debug;
use std::iter; use std::iter;
use thiserror::Error; use thiserror::Error;
@@ -172,6 +176,21 @@ pub enum InterpreterError {
FuelExhausted, FuelExhausted,
} }
pub type HeapBacking = Vec<u8>;
/// Represents a registered heap with an interpreter.
#[derive(Debug, Clone, Copy, PartialEq)]
pub struct HeapId(u32);
/// Options for initializing a heap memory region
#[derive(Debug)]
pub enum HeapInit {
/// A zero initialized heap with `size` bytes
Zeroed(usize),
/// Initializes the heap with the backing memory unchanged.
FromBacking(HeapBacking),
}
/// Maintains the [Interpreter]'s state, implementing the [State] trait. /// Maintains the [Interpreter]'s state, implementing the [State] trait.
pub struct InterpreterState<'a> { pub struct InterpreterState<'a> {
pub functions: FunctionStore<'a>, pub functions: FunctionStore<'a>,
@@ -179,7 +198,7 @@ pub struct InterpreterState<'a> {
/// Number of bytes from the bottom of the stack where the current frame's stack space is /// Number of bytes from the bottom of the stack where the current frame's stack space is
pub frame_offset: usize, pub frame_offset: usize,
pub stack: Vec<u8>, pub stack: Vec<u8>,
pub heap: Vec<u8>, pub heaps: Vec<HeapBacking>,
pub iflags: HashSet<IntCC>, pub iflags: HashSet<IntCC>,
pub fflags: HashSet<FloatCC>, pub fflags: HashSet<FloatCC>,
} }
@@ -191,7 +210,7 @@ impl Default for InterpreterState<'_> {
frame_stack: vec![], frame_stack: vec![],
frame_offset: 0, frame_offset: 0,
stack: Vec::with_capacity(1024), stack: Vec::with_capacity(1024),
heap: vec![0; 1024], heaps: Vec::new(),
iflags: HashSet::new(), iflags: HashSet::new(),
fflags: HashSet::new(), fflags: HashSet::new(),
} }
@@ -203,6 +222,57 @@ impl<'a> InterpreterState<'a> {
Self { functions, ..self } Self { functions, ..self }
} }
/// Registers a static heap and returns a reference to it
///
/// This heap reference can be used to generate a heap pointer, which
/// can be used inside the interpreter to load / store values into the heap.
///
/// ```rust
/// # use cranelift_codegen::ir::types::I64;
/// # use cranelift_interpreter::interpreter::{InterpreterState, HeapInit};
/// let mut state = InterpreterState::default();
/// let heap0 = state.register_heap(HeapInit::Zeroed(1024));
///
/// let backing = Vec::from([10u8; 24]);
/// let heap1 = state.register_heap(HeapInit::FromBacking(backing));
/// ```
pub fn register_heap(&mut self, init: HeapInit) -> HeapId {
let heap_id = HeapId(self.heaps.len() as u32);
self.heaps.push(match init {
HeapInit::Zeroed(size) => iter::repeat(0).take(size).collect(),
HeapInit::FromBacking(backing) => backing,
});
heap_id
}
/// Returns a heap address that can be used inside the interpreter
///
/// ```rust
/// # use cranelift_codegen::ir::types::I64;
/// # use cranelift_interpreter::interpreter::{InterpreterState, HeapInit};
/// let mut state = InterpreterState::default();
/// let heap_id = state.register_heap(HeapInit::Zeroed(1024));
/// let heap_base = state.get_heap_address(I64, heap_id, 0);
/// let heap_bound = state.get_heap_address(I64, heap_id, 1024);
/// ```
pub fn get_heap_address(
&self,
ty: Type,
heap_id: HeapId,
offset: u64,
) -> Result<DataValue, MemoryError> {
let size = AddressSize::try_from(ty)?;
let heap_id = heap_id.0 as u64;
let addr = Address::from_parts(size, AddressRegion::Heap, heap_id, offset)?;
self.validate_address(&addr)?;
let dv = addr.try_into()?;
Ok(dv)
}
fn current_frame_mut(&mut self) -> &mut Frame<'a> { fn current_frame_mut(&mut self) -> &mut Frame<'a> {
let num_frames = self.frame_stack.len(); let num_frames = self.frame_stack.len();
match num_frames { match num_frames {
@@ -310,23 +380,54 @@ impl<'a> State<'a, DataValue> for InterpreterState<'a> {
Address::from_parts(size, AddressRegion::Stack, 0, final_offset) Address::from_parts(size, AddressRegion::Stack, 0, final_offset)
} }
fn heap_address(&self, _size: AddressSize, _offset: u64) -> Result<Address, MemoryError> { /// Builds an [Address] for the [Heap] referenced in the currently executing function.
unimplemented!() ///
/// A CLIF Heap is essentially a GlobalValue and some metadata about that memory
/// region, such as bounds. Since heaps are based on Global Values it means that
/// once that GV is resolved we can essentially end up anywhere in memory.
///
/// To build an [Address] we perform GV resolution, and try to ensure that we end up
/// in a valid region of memory.
fn heap_address(
&self,
size: AddressSize,
heap: Heap,
offset: u64,
) -> Result<Address, MemoryError> {
let heap_data = &self.get_current_function().heaps[heap];
let heap_base = self.resolve_global_value(heap_data.base)?;
let mut addr = Address::try_from(heap_base)?;
addr.size = size;
addr.offset += offset;
// After resolving the address can point anywhere, we need to check if it's
// still valid.
self.validate_address(&addr)?;
Ok(addr)
} }
fn checked_load(&self, addr: Address, ty: Type) -> Result<DataValue, MemoryError> { fn checked_load(&self, addr: Address, ty: Type) -> Result<DataValue, MemoryError> {
let load_size = ty.bytes() as usize; let load_size = ty.bytes() as usize;
let addr_start = addr.offset as usize;
let addr_end = addr_start + load_size;
let src = match addr.region { let src = match addr.region {
AddressRegion::Stack => { AddressRegion::Stack => {
let addr_start = addr.offset as usize;
let addr_end = addr_start + load_size;
if addr_end > self.stack.len() { if addr_end > self.stack.len() {
return Err(MemoryError::OutOfBoundsLoad { addr, load_size }); return Err(MemoryError::OutOfBoundsLoad { addr, load_size });
} }
&self.stack[addr_start..addr_end] &self.stack[addr_start..addr_end]
} }
AddressRegion::Heap => {
let heap_mem = match self.heaps.get(addr.entry as usize) {
Some(mem) if addr_end <= mem.len() => mem,
_ => return Err(MemoryError::OutOfBoundsLoad { addr, load_size }),
};
&heap_mem[addr_start..addr_end]
}
_ => unimplemented!(), _ => unimplemented!(),
}; };
@@ -335,28 +436,172 @@ impl<'a> State<'a, DataValue> for InterpreterState<'a> {
fn checked_store(&mut self, addr: Address, v: DataValue) -> Result<(), MemoryError> { fn checked_store(&mut self, addr: Address, v: DataValue) -> Result<(), MemoryError> {
let store_size = v.ty().bytes() as usize; let store_size = v.ty().bytes() as usize;
let addr_start = addr.offset as usize;
let addr_end = addr_start + store_size;
let dst = match addr.region { let dst = match addr.region {
AddressRegion::Stack => { AddressRegion::Stack => {
let addr_start = addr.offset as usize;
let addr_end = addr_start + store_size;
if addr_end > self.stack.len() { if addr_end > self.stack.len() {
return Err(MemoryError::OutOfBoundsStore { addr, store_size }); return Err(MemoryError::OutOfBoundsStore { addr, store_size });
} }
&mut self.stack[addr_start..addr_end] &mut self.stack[addr_start..addr_end]
} }
AddressRegion::Heap => {
let heap_mem = match self.heaps.get_mut(addr.entry as usize) {
Some(mem) if addr_end <= mem.len() => mem,
_ => return Err(MemoryError::OutOfBoundsStore { addr, store_size }),
};
&mut heap_mem[addr_start..addr_end]
}
_ => unimplemented!(), _ => unimplemented!(),
}; };
Ok(v.write_to_slice(dst)) Ok(v.write_to_slice(dst))
} }
/// Non-Recursively resolves a global value until its address is found
fn resolve_global_value(&self, gv: GlobalValue) -> Result<DataValue, MemoryError> {
// Resolving a Global Value is a "pointer" chasing operation that lends itself to
// using a recursive solution. However, resolving this in a recursive manner
// is a bad idea because its very easy to add a bunch of global values and
// blow up the call stack.
//
// Adding to the challenges of this, is that the operations possible with GlobalValues
// mean that we cannot use a simple loop to resolve each global value, we must keep
// a pending list of operations.
// These are the possible actions that we can perform
#[derive(Debug)]
enum ResolveAction {
Resolve(GlobalValue),
/// Perform an add on the current address
Add(DataValue),
/// Load From the current address and replace it with the loaded value
Load {
/// Offset added to the base pointer before doing the load.
offset: i32,
/// Type of the loaded value.
global_type: Type,
},
}
let func = self.get_current_function();
// We start with a sentinel value that will fail if we try to load / add to it
// without resolving the base GV First.
let mut current_val = DataValue::B(false);
let mut action_stack = vec![ResolveAction::Resolve(gv)];
loop {
match action_stack.pop() {
Some(ResolveAction::Resolve(gv)) => match func.global_values[gv] {
GlobalValueData::VMContext => {
// Fetch the VMContext value from the values of the first block in the function
let index = func
.signature
.params
.iter()
.enumerate()
.find(|(_, p)| p.purpose == ArgumentPurpose::VMContext)
.map(|(i, _)| i)
// This should be validated by the verifier
.expect("No VMCtx argument was found, but one is referenced");
let first_block =
func.layout.blocks().next().expect("to have a first block");
let vmctx_value = func.dfg.block_params(first_block)[index];
current_val = self.current_frame().get(vmctx_value).clone();
}
GlobalValueData::Load {
base,
offset,
global_type,
..
} => {
action_stack.push(ResolveAction::Load {
offset: offset.into(),
global_type,
});
action_stack.push(ResolveAction::Resolve(base));
}
GlobalValueData::IAddImm {
base,
offset,
global_type,
} => {
let offset: i64 = offset.into();
let dv = DataValue::int(offset as i128, global_type)
.map_err(|_| MemoryError::InvalidAddressType(global_type))?;
action_stack.push(ResolveAction::Add(dv));
action_stack.push(ResolveAction::Resolve(base));
}
GlobalValueData::Symbol { .. } => unimplemented!(),
},
Some(ResolveAction::Add(dv)) => {
current_val = current_val
.add(dv.clone())
.map_err(|_| MemoryError::InvalidAddress(dv))?;
}
Some(ResolveAction::Load {
offset,
global_type,
}) => {
let mut addr = Address::try_from(current_val)?;
// We can forego bounds checking here since its performed in `checked_load`
addr.offset += offset as u64;
current_val = self.checked_load(addr, global_type)?;
}
// We are done resolving this, return the current value
None => return Ok(current_val),
}
}
}
fn validate_address(&self, addr: &Address) -> Result<(), MemoryError> {
match addr.region {
AddressRegion::Stack => {
let stack_len = self.stack.len() as u64;
if addr.offset > stack_len {
return Err(MemoryError::InvalidEntry {
entry: addr.entry,
max: self.heaps.len() as u64,
});
}
}
AddressRegion::Heap => {
let heap_len = self
.heaps
.get(addr.entry as usize)
.ok_or_else(|| MemoryError::InvalidEntry {
entry: addr.entry,
max: self.heaps.len() as u64,
})
.map(|heap| heap.len() as u64)?;
if addr.offset > heap_len {
return Err(MemoryError::InvalidOffset {
offset: addr.offset,
max: heap_len,
});
}
}
_ => unimplemented!(),
}
Ok(())
}
} }
#[cfg(test)] #[cfg(test)]
mod tests { mod tests {
use super::*; use super::*;
use crate::step::CraneliftTrap; use crate::step::CraneliftTrap;
use cranelift_codegen::ir::types::I64;
use cranelift_codegen::ir::TrapCode; use cranelift_codegen::ir::TrapCode;
use cranelift_reader::parse_functions; use cranelift_reader::parse_functions;
@@ -720,4 +965,51 @@ mod tests {
assert_eq!(trap, CraneliftTrap::User(TrapCode::HeapOutOfBounds)); assert_eq!(trap, CraneliftTrap::User(TrapCode::HeapOutOfBounds));
} }
/// Most heap tests are in .clif files using the filetest machinery. However, this is a sanity
/// check that the heap mechanism works without the rest of the filetest infrastructure
#[test]
fn heap_sanity_test() {
let code = "
function %heap_load_store(i64 vmctx) -> b1 {
gv0 = vmctx
gv1 = load.i64 notrap aligned gv0+0
; gv2/3 do nothing, but makes sure we understand the iadd_imm mechanism
gv2 = iadd_imm.i64 gv1, 1
gv3 = iadd_imm.i64 gv2, -1
heap0 = static gv3, min 0x1000, bound 0x1_0000_0000, offset_guard 0, index_type i64
block0(v0: i64):
v1 = iconst.i64 0
v2 = iconst.i64 123
v3 = heap_addr.i64 heap0, v1, 8
store.i64 v2, v3
v4 = load.i64 v3
v5 = icmp eq v2, v4
return v5
}";
let func = parse_functions(code).unwrap().into_iter().next().unwrap();
let mut env = FunctionStore::default();
env.add(func.name.to_string(), &func);
let mut state = InterpreterState::default().with_function_store(env);
let heap0 = state.register_heap(HeapInit::Zeroed(0x1000));
let base_addr = state.get_heap_address(I64, heap0, 0).unwrap();
// Build a vmctx struct by writing the base pointer at index 0
let mut vmctx_struct = vec![0u8; 8];
base_addr.write_to_slice(&mut vmctx_struct[..]);
// This is our vmctx "heap"
let vmctx = state.register_heap(HeapInit::FromBacking(vmctx_struct));
let vmctx_addr = state.get_heap_address(I64, vmctx, 0).unwrap();
let result = Interpreter::new(state)
.call_by_name("%heap_load_store", &[vmctx_addr])
.unwrap()
.unwrap_return();
assert_eq!(result, vec![DataValue::B(true)])
}
} }

View File

@@ -3,7 +3,7 @@
use crate::address::{Address, AddressSize}; use crate::address::{Address, AddressSize};
use cranelift_codegen::data_value::DataValue; use cranelift_codegen::data_value::DataValue;
use cranelift_codegen::ir::condcodes::{FloatCC, IntCC}; use cranelift_codegen::ir::condcodes::{FloatCC, IntCC};
use cranelift_codegen::ir::{FuncRef, Function, StackSlot, Type, Value}; use cranelift_codegen::ir::{FuncRef, Function, GlobalValue, Heap, StackSlot, Type, Value};
use cranelift_entity::PrimaryMap; use cranelift_entity::PrimaryMap;
use smallvec::SmallVec; use smallvec::SmallVec;
use thiserror::Error; use thiserror::Error;
@@ -67,13 +67,25 @@ pub trait State<'a, V> {
offset: u64, offset: u64,
) -> Result<Address, MemoryError>; ) -> Result<Address, MemoryError>;
/// Computes a heap address /// Computes a heap address
fn heap_address(&self, size: AddressSize, offset: u64) -> Result<Address, MemoryError>; fn heap_address(
&self,
size: AddressSize,
heap: Heap,
offset: u64,
) -> Result<Address, MemoryError>;
/// Retrieve a value `V` from memory at the given `address`, checking if it belongs either to the /// Retrieve a value `V` from memory at the given `address`, checking if it belongs either to the
/// stack or to one of the heaps; the number of bytes loaded corresponds to the specified [Type]. /// stack or to one of the heaps; the number of bytes loaded corresponds to the specified [Type].
fn checked_load(&self, address: Address, ty: Type) -> Result<V, MemoryError>; fn checked_load(&self, address: Address, ty: Type) -> Result<V, MemoryError>;
/// Store a value `V` into memory at the given `address`, checking if it belongs either to the /// Store a value `V` into memory at the given `address`, checking if it belongs either to the
/// stack or to one of the heaps; the number of bytes stored corresponds to the specified [Type]. /// stack or to one of the heaps; the number of bytes stored corresponds to the specified [Type].
fn checked_store(&mut self, address: Address, v: V) -> Result<(), MemoryError>; fn checked_store(&mut self, address: Address, v: V) -> Result<(), MemoryError>;
/// Given a global value, compute the final value for that global value, applying all operations
/// in intermediate global values.
fn resolve_global_value(&self, gv: GlobalValue) -> Result<V, MemoryError>;
/// Checks if an address is valid and within a known region of memory
fn validate_address(&self, address: &Address) -> Result<(), MemoryError>;
} }
#[derive(Error, Debug)] #[derive(Error, Debug)]
@@ -151,7 +163,12 @@ where
unimplemented!() unimplemented!()
} }
fn heap_address(&self, _size: AddressSize, _offset: u64) -> Result<Address, MemoryError> { fn heap_address(
&self,
_size: AddressSize,
_heap: Heap,
_offset: u64,
) -> Result<Address, MemoryError> {
unimplemented!() unimplemented!()
} }
@@ -162,4 +179,12 @@ where
fn checked_store(&mut self, _addr: Address, _v: V) -> Result<(), MemoryError> { fn checked_store(&mut self, _addr: Address, _v: V) -> Result<(), MemoryError> {
unimplemented!() unimplemented!()
} }
fn resolve_global_value(&self, _gv: GlobalValue) -> Result<V, MemoryError> {
unimplemented!()
}
fn validate_address(&self, _addr: &Address) -> Result<(), MemoryError> {
unimplemented!()
}
} }

View File

@@ -7,11 +7,12 @@ use crate::value::{Value, ValueConversionKind, ValueError, ValueResult};
use cranelift_codegen::data_value::DataValue; use cranelift_codegen::data_value::DataValue;
use cranelift_codegen::ir::condcodes::{FloatCC, IntCC}; use cranelift_codegen::ir::condcodes::{FloatCC, IntCC};
use cranelift_codegen::ir::{ use cranelift_codegen::ir::{
types, Block, FuncRef, Function, InstructionData, Opcode, TrapCode, Value as ValueRef, types, Block, FuncRef, Function, InstructionData, Opcode, TrapCode, Type, Value as ValueRef,
}; };
use log::trace; use log::trace;
use smallvec::{smallvec, SmallVec}; use smallvec::{smallvec, SmallVec};
use std::convert::{TryFrom, TryInto}; use std::convert::{TryFrom, TryInto};
use std::fmt::Debug;
use std::ops::RangeFrom; use std::ops::RangeFrom;
use thiserror::Error; use thiserror::Error;
@@ -135,11 +136,11 @@ where
Err(e) => ControlFlow::Trap(CraneliftTrap::User(memerror_to_trap(e))), Err(e) => ControlFlow::Trap(CraneliftTrap::User(memerror_to_trap(e))),
}; };
let calculate_addr = |imm: V, args: SmallVec<[V; 1]>| -> ValueResult<u64> { let calculate_addr = |addr_ty: Type, imm: V, args: SmallVec<[V; 1]>| -> ValueResult<u64> {
let imm = imm.convert(ValueConversionKind::ZeroExtend(ctrl_ty))?; let imm = imm.convert(ValueConversionKind::ZeroExtend(addr_ty))?;
let args = args let args = args
.into_iter() .into_iter()
.map(|v| v.convert(ValueConversionKind::ZeroExtend(ctrl_ty))) .map(|v| v.convert(ValueConversionKind::ZeroExtend(addr_ty)))
.collect::<ValueResult<SmallVec<[V; 1]>>>()?; .collect::<ValueResult<SmallVec<[V; 1]>>>()?;
Ok(sum(imm, args)? as u64) Ok(sum(imm, args)? as u64)
@@ -315,7 +316,7 @@ where
_ => unreachable!(), _ => unreachable!(),
}; };
let addr_value = calculate_addr(imm(), args()?)?; let addr_value = calculate_addr(types::I64, imm(), args()?)?;
let loaded = assign_or_memtrap( let loaded = assign_or_memtrap(
Address::try_from(addr_value).and_then(|addr| state.checked_load(addr, load_ty)), Address::try_from(addr_value).and_then(|addr| state.checked_load(addr, load_ty)),
); );
@@ -338,7 +339,7 @@ where
_ => unreachable!(), _ => unreachable!(),
}; };
let addr_value = calculate_addr(imm(), args_range(1..)?)?; let addr_value = calculate_addr(types::I64, imm(), args_range(1..)?)?;
let reduced = if let Some(c) = kind { let reduced = if let Some(c) = kind {
arg(0)?.convert(c)? arg(0)?.convert(c)?
} else { } else {
@@ -383,7 +384,21 @@ where
Opcode::GlobalValue => unimplemented!("GlobalValue"), Opcode::GlobalValue => unimplemented!("GlobalValue"),
Opcode::SymbolValue => unimplemented!("SymbolValue"), Opcode::SymbolValue => unimplemented!("SymbolValue"),
Opcode::TlsValue => unimplemented!("TlsValue"), Opcode::TlsValue => unimplemented!("TlsValue"),
Opcode::HeapAddr => unimplemented!("HeapAddr"), Opcode::HeapAddr => {
if let InstructionData::HeapAddr { heap, .. } = inst {
let load_ty = inst_context.controlling_type().unwrap();
let offset = calculate_addr(ctrl_ty, imm(), args()?)? as u64;
assign_or_memtrap({
AddressSize::try_from(load_ty).and_then(|addr_size| {
let addr = state.heap_address(addr_size, heap, offset)?;
let dv = DataValue::try_from(addr)?;
Ok(dv.into())
})
})
} else {
unreachable!()
}
}
Opcode::GetPinnedReg => unimplemented!("GetPinnedReg"), Opcode::GetPinnedReg => unimplemented!("GetPinnedReg"),
Opcode::SetPinnedReg => unimplemented!("SetPinnedReg"), Opcode::SetPinnedReg => unimplemented!("SetPinnedReg"),
Opcode::TableAddr => unimplemented!("TableAddr"), Opcode::TableAddr => unimplemented!("TableAddr"),