Merge pull request #5210 from wasmerio/5196-fix-llvm-relocations-and-…

…simd-operations-on-aarch64 Fix(LLVM): Implement the missing relocations on aarch64, fix SIMD operations
wasmerio · Nov 5, 2024 · dbbbe29 · dbbbe29
2 parents 0d4db42 + 5eefba5
commit dbbbe29
Show file tree

Hide file tree

Showing 14 changed files with 196 additions and 28 deletions.
diff --git a/.github/workflows/build.yml b/.github/workflows/build.yml
@@ -119,7 +119,7 @@ jobs:
             target: aarch64-apple-darwin
             artifact_name: 'wasmer-darwin-arm64'
             use_sccache: false
-            use_llvm: false
+            use_llvm: true
             build_wasm: false
             # [todo] xdoardo: Reinstate when the code we generate for aarch64 is working correctly. 
             llvm_url: 'https://github.com/wasmerio/llvm-custom-builds/releases/download/18.x/llvm-darwin-aarch64.tar.xz'

diff --git a/.github/workflows/test.yaml b/.github/workflows/test.yaml
@@ -511,8 +511,7 @@ jobs:
             os: macos-14,
             target: aarch64-apple-darwin,
             exe: '',
-            # [todo] xdoardo: Reinstate when the code we generate for aarch64 is working correctly. 
-            # llvm_url: 'https://github.com/wasmerio/llvm-custom-builds/releases/download/18.x/llvm-darwin-aarch64.tar.xz'
+            llvm_url: 'https://github.com/wasmerio/llvm-custom-builds/releases/download/18.x/llvm-darwin-aarch64.tar.xz'
           },
           {
             build: windows-x64,
@@ -744,7 +743,6 @@ jobs:
           }
         ]
         metadata: [
-          # We cannot test on macos-arm since we don't have ARM runners
           {
             build: linux-x64,
             os: ubuntu-22.04,
@@ -758,15 +756,13 @@ jobs:
             target: x86_64-apple-darwin,
             exe: '',
             llvm_url: 'https://github.com/wasmerio/llvm-custom-builds/releases/download/18.x/llvm-darwin-amd64.tar.xz'
-
           },
           {
             build: macos-arm,
             os: macos-14,
             target: aarch64-apple-darwin,
             exe: '',
-            # [todo] xdoardo: Reinstate when the code we generate for aarch64 is working correctly. 
-            # llvm_url: 'https://github.com/wasmerio/llvm-custom-builds/releases/download/18.x/llvm-darwin-aarch64.tar.xz'
+            llvm_url: 'https://github.com/wasmerio/llvm-custom-builds/releases/download/18.x/llvm-darwin-aarch64.tar.xz'
           },
           {
             build: windows-x64,

diff --git a/Cargo.lock b/Cargo.lock
diff --git a/Makefile b/Makefile
@@ -437,7 +437,7 @@ else
 endif
 
 build-docs:
-	$(CARGO_BINARY) doc $(CARGO_TARGET_FLAG) --release $(compiler_features) --document-private-items --no-deps --workspace --exclude wasmer-c-api --locked
+	$(CARGO_BINARY) doc $(CARGO_TARGET_FLAG) --release $(compiler_features) --document-private-items --no-deps --workspace --exclude wasmer-c-api --exclude wasmer-swift --locked
 
 # The tokio crate was excluded from the docs build because the code (which is not under our control)
 # does not currently compile its docs successfully

diff --git a/examples/memory.rs b/examples/memory.rs
@@ -108,8 +108,8 @@ fn main() -> anyhow::Result<()> {
     // see how we can do that:
     println!("Growing memory...");
 
-    // 'wasm-c-api' does not support direct calls to memory.grow()
-    #[cfg(not(feature = "wasm-c-api"))]
+    // 'wamr' does not support direct calls to memory.grow()
+    #[cfg(not(feature = "wamr"))]
     {
         // Here we are requesting two more pages for our memory.
         memory.grow(&mut store, 2)?;

diff --git a/examples/table.rs b/examples/table.rs
@@ -154,7 +154,6 @@ fn main() -> anyhow::Result<()> {
 
 // This test is currently failing with:
 // not implemented: Native function definitions can't be directly called from the host yet
-#[cfg(FALSE)]
 #[test]
 fn test_table() -> anyhow::Result<()> {
     main()

diff --git a/lib/compiler-llvm/src/config.rs b/lib/compiler-llvm/src/config.rs
@@ -119,11 +119,13 @@ impl LLVM {
             //  but not in the case of Aarch64, there the ABI is slightly different
             #[allow(clippy::match_single_binding)]
             match target.triple().architecture {
+                Architecture::Aarch64(_) => wasmer_types::OperatingSystem::Darwin,
                 _ => wasmer_types::OperatingSystem::Linux,
             }
         } else {
             target.triple().operating_system
         };
+
         let binary_format = if self.is_pic {
             target.triple().binary_format
         } else {

diff --git a/lib/compiler-llvm/src/object_file.rs b/lib/compiler-llvm/src/object_file.rs
@@ -251,6 +251,32 @@ where
                     object::RelocationKind::Elf(object::elf::R_LARCH_ABS64_LO20),
                     0,
                 ) => RelocationKind::LArchAbs64Lo20,
+                (
+                    object::Architecture::Aarch64,
+                    object::RelocationKind::Elf(object::elf::R_AARCH64_ADR_PREL_LO21),
+                    0,
+                ) => RelocationKind::Aarch64AdrPrelLo21,
+                (
+                    object::Architecture::Aarch64,
+                    object::RelocationKind::Elf(object::elf::R_AARCH64_ADR_PREL_PG_HI21),
+                    0,
+                ) => RelocationKind::Aarch64AdrPrelPgHi21,
+                (
+                    object::Architecture::Aarch64,
+                    object::RelocationKind::Elf(object::elf::R_AARCH64_LDST128_ABS_LO12_NC),
+                    0,
+                ) => RelocationKind::Aarch64Ldst128AbsLo12Nc,
+                (
+                    object::Architecture::Aarch64,
+                    object::RelocationKind::Elf(object::elf::R_AARCH64_ADD_ABS_LO12_NC),
+                    0,
+                ) => RelocationKind::Aarch64AddAbsLo12Nc,
+                (
+                    object::Architecture::Aarch64,
+                    object::RelocationKind::Elf(object::elf::R_AARCH64_LDST64_ABS_LO12_NC),
+                    0,
+                ) => RelocationKind::Aarch64Ldst64AbsLo12Nc,
+
                 _ => {
                     return Err(CompileError::Codegen(format!(
                         "unknown relocation {:?}",

diff --git a/lib/compiler-llvm/src/translator/intrinsics.rs b/lib/compiler-llvm/src/translator/intrinsics.rs
@@ -1104,10 +1104,6 @@ impl<'ctx> Intrinsics<'ctx> {
         intrinsics
             .throw_trap
             .add_attribute(AttributeLoc::Function, noreturn);
-        //intrinsics
-        //    .func_ref
-        //    .add_attribute(AttributeLoc::Function, intrinsics.readonly);
-
         intrinsics
     }
 }

diff --git a/lib/compiler/src/engine/link.rs b/lib/compiler/src/engine/link.rs
@@ -146,6 +146,59 @@ fn apply_relocation(
                 | read_unaligned(reloc_address as *mut u32);
             write_unaligned(reloc_address as *mut u32, reloc_abs);
         },
+        RelocationKind::Aarch64AdrPrelPgHi21 => unsafe {
+            let (reloc_address, delta) = r.for_address(body, target_func_address as u64);
+
+            let delta = delta as isize;
+            assert!(
+                ((-1 << 32)..(1 << 32)).contains(&delta),
+                "can't generate page-relative relocation with ±4GB `adrp` instruction"
+            );
+
+            let op = read_unaligned(reloc_address as *mut u32);
+            let delta = delta >> 12;
+            let immlo = ((delta as u32) & 0b11) << 29;
+            let immhi = (((delta as u32) >> 2) & 0x7ffff) << 5;
+            let mask = !((0x7ffff << 5) | (0b11 << 29));
+            let op = (op & mask) | immlo | immhi;
+
+            write_unaligned(reloc_address as *mut u32, op);
+        },
+        RelocationKind::Aarch64AdrPrelLo21 => unsafe {
+            let (reloc_address, delta) = r.for_address(body, target_func_address as u64);
+
+            let delta = delta as isize;
+            assert!(
+                ((-1 << 20)..(1 << 20)).contains(&delta),
+                "can't generate an ADR_PREL_LO21 relocation with an immediate larger than 20 bits"
+            );
+
+            let op = read_unaligned(reloc_address as *mut u32);
+            let immlo = ((delta as u32) & 0b11) << 29;
+            let immhi = (((delta as u32) >> 2) & 0x7ffff) << 5;
+            let mask = !((0x7ffff << 5) | (0b11 << 29));
+            let op = (op & mask) | immlo | immhi;
+
+            write_unaligned(reloc_address as *mut u32, op);
+        },
+        RelocationKind::Aarch64AddAbsLo12Nc => unsafe {
+            let (reloc_address, reloc_delta) = r.for_address(body, target_func_address as u64);
+            let reloc_delta = (reloc_delta as u32 & 0xfff)
+                | (read_unaligned(reloc_address as *mut u32) & 0xFFC003FF);
+            write_unaligned(reloc_address as *mut u32, reloc_delta);
+        },
+        RelocationKind::Aarch64Ldst128AbsLo12Nc => unsafe {
+            let (reloc_address, reloc_delta) = r.for_address(body, target_func_address as u64);
+            let reloc_delta = ((reloc_delta as u32 & 0xfff) >> 4) << 10
+                | (read_unaligned(reloc_address as *mut u32) & 0xFFC003FF);
+            write_unaligned(reloc_address as *mut u32, reloc_delta);
+        },
+        RelocationKind::Aarch64Ldst64AbsLo12Nc => unsafe {
+            let (reloc_address, reloc_delta) = r.for_address(body, target_func_address as u64);
+            let reloc_delta = ((reloc_delta as u32 & 0xfff) >> 3) << 10
+                | (read_unaligned(reloc_address as *mut u32) & 0xFFC003FF);
+            write_unaligned(reloc_address as *mut u32, reloc_delta);
+        },
         kind => panic!(
             "Relocation kind unsupported in the current architecture {}",
             kind

diff --git a/lib/types/src/compilation/relocation.rs b/lib/types/src/compilation/relocation.rs
@@ -46,6 +46,22 @@ pub enum RelocationKind {
     X86CallPLTRel4,
     /// x86 GOT PC-relative 4-byte
     X86GOTPCRel4,
+
+    /// R_AARCH64_ADR_PREL_LO21
+    Aarch64AdrPrelLo21,
+
+    /// R_AARCH64_ADR_PREL_PG_HI21
+    Aarch64AdrPrelPgHi21,
+
+    /// R_AARCH64_ADD_ABS_LO12_NC
+    Aarch64AddAbsLo12Nc,
+
+    /// R_AARCH64_LDST128_ABS_LO12_NC
+    Aarch64Ldst128AbsLo12Nc,
+
+    /// R_AARCH64_LDST64_ABS_LO12_NC
+    Aarch64Ldst64AbsLo12Nc,
+
     /// Arm32 call target
     Arm32Call,
     /// Arm64 call target
@@ -102,6 +118,11 @@ impl fmt::Display for RelocationKind {
             Self::LArchAbsLo12 => write!(f, "LArchAbsLo12"),
             Self::LArchAbs64Hi12 => write!(f, "LArchAbs64Hi12"),
             Self::LArchAbs64Lo20 => write!(f, "LArchAbs64Lo20"),
+            Self::Aarch64AdrPrelLo21 => write!(f, "Aarch64AdrPrelLo21"),
+            Self::Aarch64AdrPrelPgHi21 => write!(f, "Aarch64AdrPrelPgHi21"),
+            Self::Aarch64AddAbsLo12Nc => write!(f, "Aarch64AddAbsLo12Nc"),
+            Self::Aarch64Ldst128AbsLo12Nc => write!(f, "Aarch64Ldst128AbsLo12Nc"),
+            Self::Aarch64Ldst64AbsLo12Nc => write!(f, "Aarch64Ldst64AbsLo12Nc"),
             // Self::MachOX86_64Tlv => write!(f, "MachOX86_64Tlv"),
         }
     }
@@ -142,7 +163,10 @@ pub trait RelocationLike {
             | RelocationKind::Arm64Movw1
             | RelocationKind::Arm64Movw2
             | RelocationKind::Arm64Movw3
-            | RelocationKind::RiscvPCRelLo12I => {
+            | RelocationKind::RiscvPCRelLo12I
+            | RelocationKind::Aarch64AddAbsLo12Nc
+            | RelocationKind::Aarch64Ldst128AbsLo12Nc
+            | RelocationKind::Aarch64Ldst64AbsLo12Nc => {
                 let reloc_address = start + self.offset() as usize;
                 let reloc_addend = self.addend() as isize;
                 let reloc_abs = target_func_address
@@ -178,14 +202,23 @@ pub trait RelocationLike {
             }
             RelocationKind::Arm64Call
             | RelocationKind::RiscvCall
-            | RelocationKind::RiscvPCRelHi20 => {
+            | RelocationKind::RiscvPCRelHi20
+            | RelocationKind::Aarch64AdrPrelLo21 => {
                 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::Aarch64AdrPrelPgHi21 => {
+                let reloc_address = start + self.offset() as usize;
+                let reloc_addend = self.addend() as isize;
+                let target_page =
+                    (target_func_address.wrapping_add(reloc_addend as u64) & !(0xFFF)) as usize;
+                let pc_page = reloc_address & !(0xFFF);
+                (reloc_address, target_page.wrapping_sub(pc_page) as u64)
+            }
             _ => panic!("Relocation kind unsupported"),
         }
     }

diff --git a/lib/wasix/Cargo.toml b/lib/wasix/Cargo.toml
@@ -164,6 +164,8 @@ tokio = { workspace = true, features = [
 pretty_assertions.workspace = true
 tracing-test = "0.2.4"
 wasm-bindgen-test = "0.3.0"
+env_logger = { version = "0.11.5", default-features = false}
+log = "0.4.22"
 
 [target.'cfg(target_arch = "wasm32")'.dev-dependencies]
 wasm-bindgen-test = "0.3.0"