Merge #2599

2599: Feat llvm aarch64 relocs r=syrusakbary a=ptitSeb # Description Improved Aarch64 support with LLVM by adding some missing relocations. Linux/Aarch64 works fine with that, including with Universal engine. macOS/Aarch64 still have issues with Universal engine. Co-authored-by: ptitSeb <[email protected]> Co-authored-by: Syrus Akbary <[email protected]>
wasmerio · Oct 19, 2021 · 6ec1571 · 6ec1571
2 parents 18a2ca2 + b39a0cf
commit 6ec1571
Show file tree

Hide file tree

Showing 12 changed files with 263 additions and 16 deletions.
diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -22,6 +22,7 @@ Looking for changes that affect our C API? See the [C API Changelog](lib/c-api/C
 - [#2478](https://github.com/wasmerio/wasmer/pull/2478) Rename `wasm_instance_new()`’s “traps” argument to “trap”.
 
 ### Fixed
+- [#2599](https://github.com/wasmerio/wasmer/pull/2599) Fixed Universal engine for Linux/Aarch64 target.
 - [#2587](https://github.com/wasmerio/wasmer/pull/2587) Fixed deriving `WasmerEnv` when aliasing `Result`.
 - [#2518](https://github.com/wasmerio/wasmer/pull/2518) Remove temporary file used to creating an artifact when creating a Dylib engine artifact.
 - [#2494](https://github.com/wasmerio/wasmer/pull/2494) Fixed `WasmerEnv` access when using `call_indirect` with the Singlepass compiler.

diff --git a/lib/compiler-cranelift/src/compiler.rs b/lib/compiler-cranelift/src/compiler.rs
@@ -294,6 +294,7 @@ impl Compiler for CraneliftCompiler {
             function_call_trampolines,
             dynamic_function_trampolines,
             dwarf,
+            None,
         ))
     }
 }
diff --git a/lib/compiler-llvm/src/compiler.rs b/lib/compiler-llvm/src/compiler.rs
@@ -12,9 +12,10 @@ use rayon::iter::ParallelBridge;
 use rayon::prelude::{IntoParallelIterator, IntoParallelRefIterator, ParallelIterator};
 use std::sync::Arc;
 use wasmer_compiler::{
-    Compilation, CompileError, CompileModuleInfo, Compiler, CustomSection, CustomSectionProtection,
-    Dwarf, FunctionBodyData, ModuleMiddleware, ModuleTranslationState, RelocationTarget,
-    SectionBody, SectionIndex, Symbol, SymbolRegistry, Target,
+    Architecture, Compilation, CompileError, CompileModuleInfo, Compiler, CustomSection,
+    CustomSectionProtection, Dwarf, FunctionBodyData, ModuleMiddleware, ModuleTranslationState,
+    RelocationTarget, SectionBody, SectionIndex, Symbol, SymbolRegistry, Target,
+    TrampolinesSection,
 };
 use wasmer_types::entity::{EntityRef, PrimaryMap};
 use wasmer_types::{FunctionIndex, LocalFunctionIndex, SignatureIndex};
@@ -303,6 +304,37 @@ impl Compiler for LLVMCompiler {
             })
             .collect::<PrimaryMap<LocalFunctionIndex, _>>();
 
+        let trampolines = match target.triple().architecture {
+            Architecture::Aarch64(_) => {
+                let nj = 16;
+                // We create a jump to an absolute 64bits address
+                // using x17 as a scratch register, SystemV declare both x16 and x17 as Intra-Procedural  scratch register
+                // but Apple ask to just not use x16
+                // LDR x17, #8      51 00 00 58
+                // BR x17           20 02 1f d6
+                // JMPADDR          00 00 00 00 00 00 00 00
+                let onejump = [
+                    0x51, 0x00, 0x00, 0x58, 0x20, 0x02, 0x1f, 0xd6, 0, 0, 0, 0, 0, 0, 0, 0,
+                ];
+                let trampolines = Some(TrampolinesSection::new(
+                    SectionIndex::from_u32(module_custom_sections.len() as u32),
+                    nj,
+                    onejump.len(),
+                ));
+                let mut alljmps = vec![];
+                for _ in 0..nj {
+                    alljmps.extend(onejump.iter().copied());
+                }
+                module_custom_sections.push(CustomSection {
+                    protection: CustomSectionProtection::ReadExecute,
+                    bytes: SectionBody::new_with_vec(alljmps),
+                    relocations: vec![],
+                });
+                trampolines
+            }
+            _ => None,
+        };
+
         let dwarf = if !frame_section_bytes.is_empty() {
             let dwarf = Some(Dwarf::new(SectionIndex::from_u32(
                 module_custom_sections.len() as u32,
@@ -367,6 +399,7 @@ impl Compiler for LLVMCompiler {
             function_call_trampolines,
             dynamic_function_trampolines,
             dwarf,
+            trampolines,
         ))
     }
 }
diff --git a/lib/compiler-llvm/src/config.rs b/lib/compiler-llvm/src/config.rs
@@ -102,11 +102,6 @@ impl LLVM {
     fn target_triple(&self, target: &Target) -> TargetTriple {
         // Hack: we're using is_pic to determine whether this is a native
         // build or not.
-        let binary_format = if self.is_pic {
-            target.triple().binary_format
-        } else {
-            target_lexicon::BinaryFormat::Elf
-        };
         let operating_system = if target.triple().operating_system
             == wasmer_compiler::OperatingSystem::Darwin
             && !self.is_pic
@@ -117,10 +112,18 @@ impl LLVM {
             // MachO, they check whether the OS is set to Darwin.
             //
             // Since both linux and darwin use SysV ABI, this should work.
-            wasmer_compiler::OperatingSystem::Linux
+            //  but not in the case of Aarch64, there the ABI is slightly different
+            match target.triple().architecture {
+                _ => wasmer_compiler::OperatingSystem::Linux,
+            }
         } else {
             target.triple().operating_system
         };
+        let binary_format = if self.is_pic {
+            target.triple().binary_format
+        } else {
+            target_lexicon::BinaryFormat::Elf
+        };
         let triple = Triple {
             architecture: target.triple().architecture,
             vendor: target.triple().vendor.clone(),

diff --git a/lib/compiler-llvm/src/object_file.rs b/lib/compiler-llvm/src/object_file.rs
@@ -168,6 +168,19 @@ where
                 (object::RelocationKind::Elf(object::elf::R_X86_64_PC64), 0) => {
                     RelocationKind::X86PCRel8
                 }
+                (object::RelocationKind::Elf(object::elf::R_AARCH64_MOVW_UABS_G0_NC), 0) => {
+                    RelocationKind::Arm64Movw0
+                }
+                (object::RelocationKind::Elf(object::elf::R_AARCH64_MOVW_UABS_G1_NC), 0) => {
+                    RelocationKind::Arm64Movw1
+                }
+                (object::RelocationKind::Elf(object::elf::R_AARCH64_MOVW_UABS_G2_NC), 0) => {
+                    RelocationKind::Arm64Movw2
+                }
+                (object::RelocationKind::Elf(object::elf::R_AARCH64_MOVW_UABS_G3), 0) => {
+                    RelocationKind::Arm64Movw3
+                }
+                (object::RelocationKind::PltRelative, 26) => RelocationKind::Arm64Call,
                 _ => {
                     return Err(CompileError::Codegen(format!(
                         "unknown relocation {:?}",

diff --git a/lib/compiler-singlepass/src/compiler.rs b/lib/compiler-singlepass/src/compiler.rs
@@ -167,6 +167,7 @@ impl Compiler for SinglepassCompiler {
             function_call_trampolines,
             dynamic_function_trampolines,
             None,
+            None,
         ))
     }
 }

diff --git a/lib/compiler/src/function.rs b/lib/compiler/src/function.rs
@@ -109,6 +109,33 @@ impl Dwarf {
     }
 }
 
+/// Trampolines section used by ARM short jump (26bits)
+#[cfg_attr(feature = "enable-serde", derive(Deserialize, Serialize))]
+#[cfg_attr(
+    feature = "enable-rkyv",
+    derive(RkyvSerialize, RkyvDeserialize, Archive)
+)]
+#[derive(Debug, PartialEq, Eq, Clone, MemoryUsage)]
+pub struct TrampolinesSection {
+    /// SectionIndex for the actual Trampolines code
+    pub section_index: SectionIndex,
+    /// Number of jump slots in the section
+    pub slots: usize,
+    /// Slot size
+    pub size: usize,
+}
+
+impl TrampolinesSection {
+    /// Creates a `Trampolines` struct with the indice for its section, and number of slots and size of slot
+    pub fn new(section_index: SectionIndex, slots: usize, size: usize) -> Self {
+        Self {
+            section_index,
+            slots,
+            size,
+        }
+    }
+}
+
 /// The result of compiling a WebAssembly module's functions.
 #[cfg_attr(feature = "enable-serde", derive(Deserialize, Serialize))]
 #[derive(Debug, PartialEq, Eq)]
@@ -155,6 +182,9 @@ pub struct Compilation {
 
     /// Section ids corresponding to the Dwarf debug info
     debug: Option<Dwarf>,
+
+    /// Trampolines for the arch that needs it
+    trampolines: Option<TrampolinesSection>,
 }
 
 impl Compilation {
@@ -165,13 +195,15 @@ impl Compilation {
         function_call_trampolines: PrimaryMap<SignatureIndex, FunctionBody>,
         dynamic_function_trampolines: PrimaryMap<FunctionIndex, FunctionBody>,
         debug: Option<Dwarf>,
+        trampolines: Option<TrampolinesSection>,
     ) -> Self {
         Self {
             functions,
             custom_sections,
             function_call_trampolines,
             dynamic_function_trampolines,
             debug,
+            trampolines,
         }
     }
 
@@ -249,6 +281,11 @@ impl Compilation {
     pub fn get_debug(&self) -> Option<Dwarf> {
         self.debug.clone()
     }
+
+    /// Returns the Trampilines info.
+    pub fn get_trampolines(&self) -> Option<TrampolinesSection> {
+        self.trampolines.clone()
+    }
 }
 
 impl<'a> IntoIterator for &'a Compilation {

diff --git a/lib/compiler/src/lib.rs b/lib/compiler/src/lib.rs
@@ -74,7 +74,7 @@ pub use crate::error::{
 };
 pub use crate::function::{
     Compilation, CompiledFunction, CompiledFunctionFrameInfo, CustomSections, Dwarf, FunctionBody,
-    Functions,
+    Functions, TrampolinesSection,
 };
 pub use crate::jump_table::{JumpTable, JumpTableOffsets};
 pub use crate::module::CompileModuleInfo;

diff --git a/lib/compiler/src/relocation.rs b/lib/compiler/src/relocation.rs
@@ -50,6 +50,14 @@ pub enum RelocationKind {
     Arm32Call,
     /// Arm64 call target
     Arm64Call,
+    /// Arm64 movk/z part 0
+    Arm64Movw0,
+    /// Arm64 movk/z part 1
+    Arm64Movw1,
+    /// Arm64 movk/z part 2
+    Arm64Movw2,
+    /// Arm64 movk/z part 3
+    Arm64Movw3,
     // /// RISC-V call target
     // RiscvCall,
     /// Elf x86_64 32 bit signed PC relative offset to two GOT entries for GD symbol.
@@ -72,6 +80,10 @@ impl fmt::Display for RelocationKind {
             Self::X86CallPLTRel4 => write!(f, "CallPLTRel4"),
             Self::X86GOTPCRel4 => write!(f, "GOTPCRel4"),
             Self::Arm32Call | Self::Arm64Call => write!(f, "Call"),
+            Self::Arm64Movw0 => write!(f, "Arm64MovwG0"),
+            Self::Arm64Movw1 => write!(f, "Arm64MovwG1"),
+            Self::Arm64Movw2 => write!(f, "Arm64MovwG2"),
+            Self::Arm64Movw3 => write!(f, "Arm64MovwG3"),
             Self::ElfX86_64TlsGd => write!(f, "ElfX86_64TlsGd"),
             // Self::MachOX86_64Tlv => write!(f, "MachOX86_64Tlv"),
         }
@@ -121,7 +133,11 @@ impl Relocation {
     /// The function returns the relocation address and the delta.
     pub fn for_address(&self, start: usize, target_func_address: u64) -> (usize, u64) {
         match self.kind {
-            RelocationKind::Abs8 => {
+            RelocationKind::Abs8
+            | RelocationKind::Arm64Movw0
+            | RelocationKind::Arm64Movw1
+            | RelocationKind::Arm64Movw2
+            | RelocationKind::Arm64Movw3 => {
                 let reloc_address = start + self.offset as usize;
                 let reloc_addend = self.addend as isize;
                 let reloc_abs = target_func_address
@@ -155,6 +171,14 @@ impl Relocation {
                     .wrapping_add(reloc_addend as u32);
                 (reloc_address, reloc_delta_u32 as u64)
             }
+            RelocationKind::Arm64Call => {
+                let reloc_address = start + self.offset as usize;
+                let reloc_addend = self.addend as isize;
+                let reloc_delta_u32 = target_func_address
+                    .wrapping_sub(reloc_address as u64)
+                    .wrapping_add(reloc_addend as u64);
+                (reloc_address, reloc_delta_u32)
+            }
             // RelocationKind::X86PCRelRodata4 => {
             //     (start, target_func_address)
             // }

diff --git a/lib/engine-universal/src/artifact.rs b/lib/engine-universal/src/artifact.rs
@@ -122,6 +122,7 @@ impl UniversalArtifact {
             custom_sections: compilation.get_custom_sections(),
             custom_section_relocations: compilation.get_custom_section_relocations(),
             debug: compilation.get_debug(),
+            trampolines: compilation.get_trampolines(),
         };
         let serializable = SerializableModule {
             compilation: serializable_compilation,
@@ -194,6 +195,7 @@ impl UniversalArtifact {
             serializable.compilation.function_relocations.clone(),
             &custom_sections,
             &serializable.compilation.custom_section_relocations,
+            &serializable.compilation.trampolines,
         );
 
         // Compute indices into the shared signature table.
@@ -221,6 +223,7 @@ impl UniversalArtifact {
             }
             None => None,
         };
+
         // Make all code compiled thus far executable.
         inner_engine.publish_compiled_code();
-Original file line number
+Diff line change
@@ Expand Up / @@ -294,6 +294,7 @@ impl Compiler for CraneliftCompiler { @@
                 function_call_trampolines,
                 dynamic_function_trampolines,
                 dwarf,
+                None,
             ))
         }
     }