T0b1/wasmtime
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Record information about sections of emitted code+data.

Browse Source 
The result of the emitter is a vector of bytes holding machine code,
jump tables, and (in the future) other read-only data.  Some clients,
notably Firefox's Wasm compiler, needs to separate the machine code
from the data in order to insert more code directly after the code
generated by Cranelift.

To make such separation possible, we record more information about the
emitted bytes: the sizes of each of the sections of code, jump tables,
and read-only data, as well as the locations within the code that
reference (PC-relatively) the jump tables and read-only data.
This commit is contained in:
Lars T Hansen
2019-05-21 15:53:46 +02:00
committed by Lars T Hansen
parent 70f79d23bf
commit 420850adf0
12 changed files with 166 additions and 73 deletions
Download Patch File Download Diff File Expand all files Collapse all files

57
cranelift/codegen/src/binemit/memorysink.rs
View File

@@ -14,7 +14,7 @@
//! relocations to a `RelocSink` trait object. Relocations are less frequent than the
//! `CodeSink::put*` methods, so the performance impact of the virtual callbacks is less severe.
use super::{Addend, CodeOffset, CodeSink, Reloc};
use super::{Addend, CodeInfo, CodeOffset, CodeSink, Reloc};
use crate::ir::{ExternalName, JumpTable, SourceLoc, TrapCode};
use core::ptr::write_unaligned;
@@ -30,12 +30,14 @@ use core::ptr::write_unaligned;
/// Note that `MemoryCodeSink` writes multi-byte values in the native byte order of the host. This
/// is not the right thing to do for cross compilation.
pub struct MemoryCodeSink<'a> {
/// Pointer to start of sink's preallocated memory.
data: *mut u8,
/// Offset is isize because its major consumer needs it in that form.
offset: isize,
/// Size of the machine code portion of output
pub code_size: isize,
relocs: &'a mut RelocSink,
traps: &'a mut TrapSink,
/// Information about the generated code and read-only data.
pub info: CodeInfo,
}
impl<'a> MemoryCodeSink<'a> {
@@ -47,7 +49,12 @@ impl<'a> MemoryCodeSink<'a> {
Self {
data,
offset: 0,
info: CodeInfo {
code_size: 0,
jumptables_size: 0,
rodata_size: 0,
total_size: 0,
},
relocs,
traps,
}
@@ -75,40 +82,35 @@ pub trait TrapSink {
fn trap(&mut self, _: CodeOffset, _: SourceLoc, _: TrapCode);
}
impl<'a> MemoryCodeSink<'a> {
fn write<T>(&mut self, x: T) {
unsafe {
#[cfg_attr(feature = "cargo-clippy", allow(clippy::cast_ptr_alignment))]
write_unaligned(self.data.offset(self.offset) as *mut T, x);
self.offset += std::mem::size_of::<T>() as isize;
}
}
}
impl<'a> CodeSink for MemoryCodeSink<'a> {
fn offset(&self) -> CodeOffset {
self.offset as CodeOffset
}
fn put1(&mut self, x: u8) {
unsafe {
write_unaligned(self.data.offset(self.offset), x);
}
self.offset += 1;
self.write(x);
}
fn put2(&mut self, x: u16) {
unsafe {
#[cfg_attr(feature = "cargo-clippy", allow(clippy::cast_ptr_alignment))]
write_unaligned(self.data.offset(self.offset) as *mut u16, x);
}
self.offset += 2;
self.write(x);
}
fn put4(&mut self, x: u32) {
unsafe {
#[cfg_attr(feature = "cargo-clippy", allow(clippy::cast_ptr_alignment))]
write_unaligned(self.data.offset(self.offset) as *mut u32, x);
}
self.offset += 4;
self.write(x);
}
fn put8(&mut self, x: u64) {
unsafe {
#[cfg_attr(feature = "cargo-clippy", allow(clippy::cast_ptr_alignment))]
write_unaligned(self.data.offset(self.offset) as *mut u64, x);
}
self.offset += 8;
self.write(x);
}
fn reloc_ebb(&mut self, rel: Reloc, ebb_offset: CodeOffset) {
@@ -131,8 +133,17 @@ impl<'a> CodeSink for MemoryCodeSink<'a> {
self.traps.trap(ofs, srcloc, code);
}
fn begin_jumptables(&mut self) {
self.info.code_size = self.offset();
}
fn begin_rodata(&mut self) {
self.code_size = self.offset;
self.info.jumptables_size = self.offset() - self.info.code_size;
}
fn end_codegen(&mut self) {
self.info.rodata_size = self.offset() - self.info.jumptables_size;
self.info.total_size = self.offset();
}
}

50
cranelift/codegen/src/binemit/mod.rs
View File

@@ -33,6 +33,8 @@ pub enum Reloc {
Abs8,
/// x86 PC-relative 4-byte
X86PCRel4,
/// x86 PC-relative 4-byte offset to trailing rodata
X86PCRelRodata4,
/// x86 call to PC-relative 4-byte
X86CallPCRel4,
/// x86 call to PLT-relative 4-byte
@@ -55,6 +57,7 @@ impl fmt::Display for Reloc {
Reloc::Abs4 => write!(f, "Abs4"),
Reloc::Abs8 => write!(f, "Abs8"),
Reloc::X86PCRel4 => write!(f, "PCRel4"),
Reloc::X86PCRelRodata4 => write!(f, "PCRelRodata4"),
Reloc::X86CallPCRel4 => write!(f, "CallPCRel4"),
Reloc::X86CallPLTRel4 => write!(f, "CallPLTRel4"),
Reloc::X86GOTPCRel4 => write!(f, "GOTPCRel4"),
@@ -63,6 +66,38 @@ impl fmt::Display for Reloc {
}
}
/// Container for information about a vector of compiled code and its supporting read-only data.
///
/// The code starts at offset 0 and is followed optionally by relocatable jump tables and copyable
/// (raw binary) read-only data. Any padding between sections is always part of the section that
/// precedes the boundary between the sections.
#[derive(PartialEq)]
pub struct CodeInfo {
/// Number of bytes of machine code (the code starts at offset 0).
pub code_size: CodeOffset,
/// Number of bytes of jumptables.
pub jumptables_size: CodeOffset,
/// Number of bytes of rodata.
pub rodata_size: CodeOffset,
/// Number of bytes in total.
pub total_size: CodeOffset,
}
impl CodeInfo {
/// Offset of any relocatable jump tables, or equal to rodata if there are no jump tables.
pub fn jumptables(&self) -> CodeOffset {
self.code_size
}
/// Offset of any copyable read-only data, or equal to total_size if there are no rodata.
pub fn rodata(&self) -> CodeOffset {
self.code_size + self.jumptables_size
}
}
/// Abstract interface for adding bytes to the code segment.
///
/// A `CodeSink` will receive all of the machine code for a function. It also accepts relocations
@@ -95,8 +130,14 @@ pub trait CodeSink {
/// Add trap information for the current offset.
fn trap(&mut self, _: TrapCode, _: SourceLoc);
/// Code output is complete, read-only data may follow.
/// Machine code output is complete, jump table data may follow.
fn begin_jumptables(&mut self);
/// Jump table output is complete, raw read-only data may follow.
fn begin_rodata(&mut self);
/// Read-only data output is complete, we're done.
fn end_codegen(&mut self);
}
/// Report a bad encoding error.
@@ -127,7 +168,7 @@ where
}
}
sink.begin_rodata();
sink.begin_jumptables();
// output jump tables
for (jt, jt_data) in func.jump_tables.iter() {
@@ -137,4 +178,9 @@ where
sink.put4(rel_offset as u32)
}
}
sink.begin_rodata();
// TODO: No read-only data (constant pools) at this time.
sink.end_codegen();
}

21
cranelift/codegen/src/binemit/relaxation.rs
View File

@@ -27,7 +27,7 @@
//! ebb23:
//! ```
use crate::binemit::CodeOffset;
use crate::binemit::{CodeInfo, CodeOffset};
use crate::cursor::{Cursor, FuncCursor};
use crate::ir::{Function, InstructionData, Opcode};
use crate::isa::{EncInfo, TargetIsa};
@@ -40,7 +40,7 @@ use log::debug;
/// Relax branches and compute the final layout of EBB headers in `func`.
///
/// Fill in the `func.offsets` table so the function is ready for binary emission.
pub fn relax_branches(func: &mut Function, isa: &TargetIsa) -> CodegenResult<CodeOffset> {
pub fn relax_branches(func: &mut Function, isa: &TargetIsa) -> CodegenResult<CodeInfo> {
let _tt = timing::relax_branches();
let encinfo = isa.encoding_info();
@@ -109,6 +109,9 @@ pub fn relax_branches(func: &mut Function, isa: &TargetIsa) -> CodegenResult<Cod
}
}
let code_size = offset;
let jumptables = offset;
for (jt, jt_data) in func.jump_tables.iter() {
func.jt_offsets[jt] = offset;
// TODO: this should be computed based on the min size needed to hold
@@ -116,7 +119,19 @@ pub fn relax_branches(func: &mut Function, isa: &TargetIsa) -> CodegenResult<Cod
offset += jt_data.len() as u32 * 4;
}
Ok(offset)
let jumptables_size = offset - jumptables;
let rodata = offset;
// TODO: Once we have constant pools we'll do some processing here to update offset.
let rodata_size = offset - rodata;
Ok(CodeInfo {
code_size,
jumptables_size,
rodata_size,
total_size: offset,
})
}
/// Convert `jump` instructions to `fallthrough` instructions where possible and verify that any

34
cranelift/codegen/src/context.rs
View File

@@ -10,7 +10,7 @@
//! single ISA instance.
use crate::binemit::{
relax_branches, shrink_instructions, CodeOffset, MemoryCodeSink, RelocSink, TrapSink,
relax_branches, shrink_instructions, CodeInfo, MemoryCodeSink, RelocSink, TrapSink,
};
use crate::dce::do_dce;
use crate::dominator_tree::DominatorTree;
@@ -93,20 +93,21 @@ impl Context {
/// This function calls `compile` and `emit_to_memory`, taking care to resize `mem` as
/// needed, so it provides a safe interface.
///
/// Returns the size of the function's code and the size of the read-only data.
/// Returns information about the function's code and read-only data.
pub fn compile_and_emit(
&mut self,
isa: &TargetIsa,
mem: &mut Vec<u8>,
relocs: &mut RelocSink,
traps: &mut TrapSink,
) -> CodegenResult<(CodeOffset, CodeOffset)> {
let total_size = self.compile(isa)?;
) -> CodegenResult<CodeInfo> {
let info = self.compile(isa)?;
let old_len = mem.len();
mem.resize(old_len + total_size as usize, 0);
let code_size =
mem.resize(old_len + info.total_size as usize, 0);
let new_info =
unsafe { self.emit_to_memory(isa, mem.as_mut_ptr().add(old_len), relocs, traps) };
Ok((code_size, total_size - code_size))
debug_assert!(new_info == info);
Ok(info)
}
/// Compile the function.
@@ -115,8 +116,8 @@ impl Context {
/// represented by `isa`. This does not include the final step of emitting machine code into a
/// code sink.
///
/// Returns the size of the function's code and read-only data.
pub fn compile(&mut self, isa: &TargetIsa) -> CodegenResult<CodeOffset> {
/// Returns information about the function's code and read-only data.
pub fn compile(&mut self, isa: &TargetIsa) -> CodegenResult<CodeInfo> {
let _tt = timing::compile();
self.verify_if(isa)?;
@@ -160,18 +161,18 @@ impl Context {
/// This function is unsafe since it does not perform bounds checking on the memory buffer,
/// and it can't guarantee that the `mem` pointer is valid.
///
/// Returns the size of the function's code.
/// Returns information about the emitted code and data.
pub unsafe fn emit_to_memory(
&self,
isa: &TargetIsa,
mem: *mut u8,
relocs: &mut RelocSink,
traps: &mut TrapSink,
) -> CodeOffset {
) -> CodeInfo {
let _tt = timing::binemit();
let mut sink = MemoryCodeSink::new(mem, relocs, traps);
isa.emit_function_to_memory(&self.func, &mut sink);
sink.code_size as CodeOffset
sink.info
}
/// Run the verifier on the function.
@@ -325,12 +326,13 @@ impl Context {
Ok(())
}
/// Run the branch relaxation pass and return the final code size.
pub fn relax_branches(&mut self, isa: &TargetIsa) -> CodegenResult<CodeOffset> {
let code_size = relax_branches(&mut self.func, isa)?;
/// Run the branch relaxation pass and return information about the function's code and
/// read-only data.
pub fn relax_branches(&mut self, isa: &TargetIsa) -> CodegenResult<CodeInfo> {
let info = relax_branches(&mut self.func, isa)?;
self.verify_if(isa)?;
self.verify_locations_if(isa)?;
Ok(code_size)
Ok(info)
}
/// Builds ranges and location for specified value labels.

1
cranelift/codegen/src/isa/x86/binemit.rs
View File

@@ -339,4 +339,5 @@ fn disp4<CS: CodeSink + ?Sized>(destination: Ebb, func: &Function, sink: &mut CS
fn jt_disp4<CS: CodeSink + ?Sized>(jt: JumpTable, func: &Function, sink: &mut CS) {
let delta = func.jt_offsets[jt].wrapping_sub(sink.offset() + 4);
sink.put4(delta);
sink.reloc_jt(Reloc::X86PCRelRodata4, jt);
}

6
cranelift/faerie/src/backend.rs
View File

@@ -138,9 +138,9 @@ impl Backend for FaerieBackend {
name: &str,
ctx: &cranelift_codegen::Context,
namespace: &ModuleNamespace<Self>,
code_size: u32,
total_size: u32,
) -> ModuleResult<FaerieCompiledFunction> {
let mut code: Vec<u8> = vec![0; code_size as usize];
let mut code: Vec<u8> = vec![0; total_size as usize];
// Non-lexical lifetimes would obviate the braces here.
{
@@ -153,7 +153,7 @@ impl Backend for FaerieBackend {
};
if let Some(ref mut trap_manifest) = self.trap_manifest {
let mut trap_sink = FaerieTrapSink::new(name, code_size);
let mut trap_sink = FaerieTrapSink::new(name, total_size);
unsafe {
ctx.emit_to_memory(
&*self.isa,

4
cranelift/filetests/filetests/isa/x86/binary64.clif
View File

@@ -1424,8 +1424,8 @@ function %I64_JT(i64 [%rdi]) {
ebb0(v0: i64 [%rdi]):
; Note: The next two lines will need to change whenever instructions are
; added or removed from this test.
[-, %rax] v1 = jump_table_base.i64 jt0 ; bin: 48 8d 05 00000039
[-, %r10] v2 = jump_table_base.i64 jt0 ; bin: 4c 8d 15 00000032
[-, %rax] v1 = jump_table_base.i64 jt0 ; bin: 48 8d 05 00000039 PCRelRodata4(jt0)
[-, %r10] v2 = jump_table_base.i64 jt0 ; bin: 4c 8d 15 00000032 PCRelRodata4(jt0)
[-, %rbx] v10 = iconst.i64 1
[-, %r13] v11 = iconst.i64 2

21
cranelift/filetests/src/test_binemit.rs
View File

@@ -5,8 +5,7 @@
use crate::match_directive::match_directive;
use crate::subtest::{Context, SubTest, SubtestResult};
use cranelift_codegen::binemit;
use cranelift_codegen::binemit::{CodeSink, RegDiversions};
use cranelift_codegen::binemit::{self, CodeInfo, CodeSink, RegDiversions};
use cranelift_codegen::dbg::DisplayList;
use cranelift_codegen::ir;
use cranelift_codegen::ir::entities::AnyEntity;
@@ -96,9 +95,12 @@ impl binemit::CodeSink for TextSink {
write!(self.text, "{} ", code).unwrap();
}
fn begin_rodata(&mut self) {
fn begin_jumptables(&mut self) {
self.code_size = self.offset
}
fn begin_rodata(&mut self) {}
fn end_codegen(&mut self) {}
}
impl SubTest for TestBinEmit {
@@ -164,7 +166,7 @@ impl SubTest for TestBinEmit {
}
// Relax branches and compute EBB offsets based on the encodings.
let code_size = binemit::relax_branches(&mut func, isa)
let CodeInfo { total_size, .. } = binemit::relax_branches(&mut func, isa)
.map_err(|e| pretty_error(&func, context.isa, e))?;
// Collect all of the 'bin:' directives on instructions.
@@ -288,7 +290,7 @@ impl SubTest for TestBinEmit {
}
}
sink.begin_rodata();
sink.begin_jumptables();
for (jt, jt_data) in func.jump_tables.iter() {
let jt_offset = func.jt_offsets[jt];
@@ -298,10 +300,15 @@ impl SubTest for TestBinEmit {
}
}
if sink.offset != code_size {
sink.begin_rodata();
// TODO: Read-only (constant pool) data.
sink.end_codegen();
if sink.offset != total_size {
return Err(format!(
"Expected code size {}, got {}",
code_size, sink.offset
total_size, sink.offset
));
}

13
cranelift/filetests/src/test_compile.rs
View File

@@ -4,8 +4,9 @@
use crate::subtest::{run_filecheck, Context, SubTest, SubtestResult};
use cranelift_codegen;
use cranelift_codegen::binemit::{self, CodeInfo};
use cranelift_codegen::ir;
use cranelift_codegen::print_errors::pretty_error;
use cranelift_codegen::{binemit, ir};
use cranelift_reader::TestCommand;
use log::info;
use std::borrow::Cow;
@@ -38,13 +39,13 @@ impl SubTest for TestCompile {
let isa = context.isa.expect("compile needs an ISA");
let mut comp_ctx = cranelift_codegen::Context::for_function(func.into_owned());
let code_size = comp_ctx
let CodeInfo { total_size, .. } = comp_ctx
.compile(isa)
.map_err(|e| pretty_error(&comp_ctx.func, context.isa, e))?;
info!(
"Generated {} bytes of code:\n{}",
code_size,
total_size,
comp_ctx.func.display(isa)
);
@@ -56,10 +57,10 @@ impl SubTest for TestCompile {
&mut sink,
);
if sink.offset != code_size {
if sink.offset != total_size {
return Err(format!(
"Expected code size {}, got {}",
code_size, sink.offset
total_size, sink.offset
));
}
@@ -105,5 +106,7 @@ impl binemit::CodeSink for SizeSink {
}
fn reloc_jt(&mut self, _reloc: binemit::Reloc, _jt: ir::JumpTable) {}
fn trap(&mut self, _code: ir::TrapCode, _srcloc: ir::SourceLoc) {}
fn begin_jumptables(&mut self) {}
fn begin_rodata(&mut self) {}
fn end_codegen(&mut self) {}
}

11
cranelift/module/src/module.rs
View File

@@ -8,8 +8,9 @@
use super::HashMap;
use crate::data_context::DataContext;
use crate::Backend;
use cranelift_codegen::binemit::{self, CodeInfo};
use cranelift_codegen::entity::{entity_impl, PrimaryMap};
use cranelift_codegen::{binemit, ir, isa, CodegenError, Context};
use cranelift_codegen::{ir, isa, CodegenError, Context};
use failure::Fail;
use log::info;
use std::borrow::ToOwned;
@@ -519,7 +520,7 @@ where
/// Define a function, producing the function body from the given `Context`.
///
/// Returns the size of the function's code.
/// Returns the size of the function's code and constant data.
///
/// Note: After calling this function the given `Context` will contain the compiled function.
pub fn define_function(
@@ -527,7 +528,7 @@ where
func: FuncId,
ctx: &mut Context,
) -> ModuleResult<binemit::CodeOffset> {
let code_size = ctx.compile(self.backend.isa()).map_err(|e| {
let CodeInfo { total_size, .. } = ctx.compile(self.backend.isa()).map_err(|e| {
info!(
"defining function {}: {}",
func,
@@ -550,12 +551,12 @@ where
&ModuleNamespace::<B> {
contents: &self.contents,
},
code_size,
total_size,
)?);
self.contents.functions[func].compiled = compiled;
self.functions_to_finalize.push(func);
Ok(code_size)
Ok(total_size)
}
/// Define a function, producing the data contents from the given `DataContext`.

11
cranelift/src/compile.rs
View File

@@ -65,7 +65,7 @@ fn handle_module(
let mut mem = vec![];
// Compile and encode the result to machine code.
let (code_size, rodata_size) = context
let code_info = context
.compile_and_emit(isa, &mut mem, &mut relocs, &mut traps)
.map_err(|err| pretty_error(&context.func, Some(isa), err))?;
@@ -74,7 +74,14 @@ fn handle_module(
}
if flag_disasm {
print_all(isa, &mem, code_size, rodata_size, &relocs, &traps)?;
print_all(
isa,
&mem,
code_info.code_size,
code_info.jumptables_size + code_info.rodata_size,
&relocs,
&traps,
)?;
}
}

8
cranelift/src/wasm.rs
View File

@@ -176,16 +176,16 @@ fn handle_module(
return Err(pretty_verifier_error(&context.func, fisa.isa, None, errors));
}
} else {
let (code_size, rodata_size) = context
let code_info = context
.compile_and_emit(isa, &mut mem, &mut relocs, &mut traps)
.map_err(|err| pretty_error(&context.func, fisa.isa, err))?;
if flag_print_size {
println!(
"Function #{} code size: {} bytes",
func_index, code_size + rodata_size
func_index, code_info.total_size,
);
total_module_code_size += code_size + rodata_size;
total_module_code_size += code_info.total_size;
println!(
"Function #{} bytecode size: {} bytes",
func_index,
@@ -194,7 +194,7 @@ fn handle_module(
}
if flag_print_disasm {
saved_sizes = Some((code_size, rodata_size));
saved_sizes = Some((code_info.code_size, code_info.jumptables_size + code_info.rodata_size));
}
}