Merge #2681

2681: fix singlepass slow compilation of many function wasm by using dynasm VecAssembler r=ptitSeb a=ailisp

<!--
Prior to submitting a PR, review the CONTRIBUTING.md document for recommendations on how to test:
https://github.com/wasmerio/wasmer/blob/master/CONTRIBUTING.md#pull-requests

-->

<!--
Provide details regarding the change including motivation,
links to related issues, and the context of the PR.
-->

We found in certain benchmark (a wasm with 150_000 no-op functions), compilation is slowed down by dynasmrt's mmapand mprotect syscalls, however, those mmap isn't used in wasmer and can be replaced by simpler VecAssembler. In this [bench](#71) it's a 10x speedup in singlepass backend compilation time.

- [ ] Add a short description of the change to the CHANGELOG.md file

Co-authored-by: Bo Yao <[email protected]>

lib/compiler-singlepass/src/codegen_x64.rs

@@ -1,7 +1,7 @@
 use crate::address_map::get_function_address_map;
 use crate::config::{Intrinsic, IntrinsicKind};
 use crate::{common_decl::*, config::Singlepass, emitter_x64::*, machine::Machine, x64_decl::*};
-use dynasmrt::{x64::Assembler, DynamicLabel};
+use dynasmrt::{x64::X64Relocation, DynamicLabel, VecAssembler};
 use memoffset::offset_of;
 use smallvec::{smallvec, SmallVec};
 use std::iter;
@@ -23,6 +23,8 @@ use wasmer_types::{
 };
 use wasmer_vm::{TableStyle, TrapCode, VMBuiltinFunctionIndex, VMOffsets};
 
+type Assembler = VecAssembler<X64Relocation>;
+
 /// The singlepass per-function code generator.
 pub struct FuncGen<'a> {
     // Immutable properties assigned at creation time.
@@ -2001,7 +2003,7 @@ impl<'a> FuncGen<'a> {
                 .collect(),
         );
 
-        let mut assembler = Assembler::new().unwrap();
+        let mut assembler = Assembler::new(0);
         let special_labels = SpecialLabelSet {
             integer_division_by_zero: assembler.get_label(),
             integer_overflow: assembler.get_label(),
@@ -8850,7 +8852,7 @@ fn sort_call_movs(movs: &mut [(Location, GPR)]) {
 
 // Standard entry trampoline.
 pub fn gen_std_trampoline(sig: &FunctionType) -> FunctionBody {
-    let mut a = Assembler::new().unwrap();
+    let mut a = Assembler::new(0);
 
     // Calculate stack offset.
     let mut stack_offset: u32 = 0;
@@ -8952,7 +8954,7 @@ pub fn gen_std_dynamic_import_trampoline(
     vmoffsets: &VMOffsets,
     sig: &FunctionType,
 ) -> FunctionBody {
-    let mut a = Assembler::new().unwrap();
+    let mut a = Assembler::new(0);
 
     // Allocate argument array.
     let stack_offset: usize = 16 * std::cmp::max(sig.params().len(), sig.results().len()) + 8; // 16 bytes each + 8 bytes sysv call padding
@@ -9046,7 +9048,7 @@ pub fn gen_import_call_trampoline(
     index: FunctionIndex,
     sig: &FunctionType,
 ) -> CustomSection {
-    let mut a = Assembler::new().unwrap();
+    let mut a = Assembler::new(0);
 
     // TODO: ARM entry trampoline is not emitted.
 

lib/compiler-singlepass/src/emitter_x64.rs

@@ -1,6 +1,10 @@
 pub use crate::x64_decl::{GPR, XMM};
 use dynasm::dynasm;
-use dynasmrt::{x64::Assembler, AssemblyOffset, DynamicLabel, DynasmApi, DynasmLabelApi};
+use dynasmrt::{
+    x64::X64Relocation, AssemblyOffset, DynamicLabel, DynasmApi, DynasmLabelApi, VecAssembler,
+};
+
+type Assembler = VecAssembler<X64Relocation>;
 
 /// Dynasm proc-macro checks for an `.arch` expression in a source file to
 /// determine the architecture it should use.

lib/compiler-singlepass/src/machine.rs

@@ -537,12 +537,14 @@ impl Machine {
 #[cfg(test)]
 mod test {
     use super::*;
-    use dynasmrt::x64::Assembler;
+    use dynasmrt::x64::X64Relocation;
+    use dynasmrt::VecAssembler;
+    type Assembler = VecAssembler<X64Relocation>;
 
     #[test]
     fn test_release_locations_keep_state_nopanic() {
         let mut machine = Machine::new();
-        let mut assembler = Assembler::new().unwrap();
+        let mut assembler = Assembler::new(0);
         let locs = machine.acquire_locations(
             &mut assembler,
             &(0..10)