feat(arch):support thumbv6m

Author: weizhiao

Cargo.lock

@@ -17,7 +17,7 @@ authors = ["wzhao <1207410841@qq.com>"]
 edition = "2024"
 repository = "https://github.com/weizhiao/elf_loader"
 license = "MIT/Apache-2.0"
-rust-version = "1.88.0"
+rust-version = "1.93.0"
 
 [workspace]
 members = ["crates/windows-elf-loader", "crates/mini-loader", "tools/gen-elf"]
@@ -73,7 +73,7 @@ optional = true
 features = ["alloc"]
 
 [dependencies.spin]
-version = "0.9"
+version = "0.10"
 default-features = false
 features = ["mutex", "rwlock", "spin_mutex"]
 
@@ -110,7 +110,11 @@ version = []
 # Enable logging.
 log = ["dep:log"]
 # support target without native pointer size atomic operation
-portable-atomic = ["dep:portable-atomic", "dep:portable-atomic-util"]
+portable-atomic = [
+	"dep:portable-atomic",
+	"dep:portable-atomic-util",
+	"spin/portable_atomic",
+]
 
 [[example]]
 name = "relocate_dylib"

Cargo.toml

@@ -30,7 +30,7 @@ pub const REL_COPY: u32 = R_ARM_COPY;
 /// TLS TPOFF relocation type - set to TLS offset relative to thread pointer.
 pub const REL_TPOFF: u32 = R_ARM_TLS_TPOFF32;
 /// TLSDESC relocation type - set to a function pointer and an argument.
-pub const REL_TLSDESC: u32 = 0;
+pub const REL_TLSDESC: u32 = 0xffff_ffff;
 
 /// Get the current thread pointer using architecture-specific register.
 #[inline(always)]
@@ -62,56 +62,44 @@ pub(crate) const RESOLVE_FUNCTION_OFFSET: usize = 2;
 /// # Safety
 /// This function uses naked assembly and must be called with the correct
 /// stack layout set up by the PLT stub code.
-#[cfg(target_feature = "vfp2")]
 #[unsafe(naked)]
 pub(crate) extern "C" fn dl_runtime_resolve() {
     core::arch::naked_asm!(
         "
         // sp has original lr (4 bytes)
         // push r0-r4 (5 regs, 20 bytes). sp aligned to 8 bytes (aligned - 24).
         push {{r0, r1, r2, r3, r4}}
-        vpush {{d0, d1, d2, d3, d4, d5, d6, d7}}
-        
+        ",
+        #[cfg(target_feature = "vfp2")]
+        "vpush {{d0, d1, d2, d3, d4, d5, d6, d7}}",
+        "
         // r0 = link_map (GOT[1])
-        ldr r0, [lr, #-4]
+        // Case for thumb-1 compatibility: mov + sub + ldr
+        mov r0, lr
+        subs r0, r0, #4
+        ldr r0, [r0]
         
         // r1 = index
-        add r1, lr, #4
-        sub r1, ip, r1
-        lsr r1, r1, #2
+        mov r1, lr
+        adds r1, r1, #4
+        mov r2, ip
+        subs r1, r2, r1
+        lsrs r1, r1, #2
         
-        blx {0}
+        bl {fixup}
         
         mov ip, r0
-        
-        vpop {{d0, d1, d2, d3, d4, d5, d6, d7}}
-        pop {{r0, r1, r2, r3, r4, lr}}
-        bx ip
         ",
-        sym crate::relocation::dl_fixup,
-    )
-}
-
-#[cfg(not(target_feature = "vfp2"))]
-#[unsafe(naked)]
-pub(crate) extern "C" fn dl_runtime_resolve() {
-    core::arch::naked_asm!(
+        #[cfg(target_feature = "vfp2")]
+        "vpop {{d0, d1, d2, d3, d4, d5, d6, d7}}",
         "
-        push {{r0, r1, r2, r3, r4}}
-        
-        ldr r0, [lr, #-4]
-        
-        add r1, lr, #4
-        sub r1, ip, r1
-        lsr r1, r1, #2
-        
-        blx {0}
-        
-        mov ip, r0
-        pop {{r0, r1, r2, r3, r4, lr}}
+        pop {{r0, r1, r2, r3, r4}}
+        // pop original lr into r2 then move to lr
+        pop {{r2}}
+        mov lr, r2
         bx ip
         ",
-        sym crate::relocation::dl_fixup,
+        fixup = sym crate::relocation::dl_fixup,
     )
 }
 

src/arch/arm.rs

@@ -64,6 +64,6 @@ pub(crate) fn prepare_lazy_bind(got: *mut usize, dylib: usize) {
     unsafe {
         got.add(DYLIB_OFFSET).write(dylib);
         got.add(RESOLVE_FUNCTION_OFFSET)
-            .write(dl_runtime_resolve as usize);
+            .write(dl_runtime_resolve as *const () as usize);
     }
 }

src/arch/mod.rs

@@ -45,107 +45,94 @@ pub(crate) unsafe fn get_thread_pointer() -> *mut u8 {
 pub(crate) const DYLIB_OFFSET: usize = 1;
 pub(crate) const RESOLVE_FUNCTION_OFFSET: usize = 0;
 
-macro_rules! riscv32_dl_runtime_resolve {
-    ($save_fprs:expr, $restore_fprs:expr) => {
-        #[unsafe(naked)]
-        pub(crate) extern "C" fn dl_runtime_resolve() {
-            core::arch::naked_asm!(
-                "
-                // 保存整数参数寄存器
-                // ra, a0-a7: 9 * 4 = 36 bytes
-                // 栈帧总大小设为 112 字节以保持 16 字节对齐
-                addi sp,sp,-112
-                sw ra,0(sp)
-                sw a0,4(sp)
-                sw a1,8(sp)
-                sw a2,12(sp)
-                sw a3,16(sp)
-                sw a4,20(sp)
-                sw a5,24(sp)
-                sw a6,28(sp)
-                sw a7,32(sp)
-                ",
-                $save_fprs,
-                "
-                // 这两个是plt代码设置的
-                mv a0,t0
-                srli a1,t1,3
-                // 调用重定位函数
-                call {0}
-                // 恢复参数寄存器
-                mv t1,a0
-                lw ra,0(sp)
-                lw a0,4(sp)
-                lw a1,8(sp)
-                lw a2,12(sp)
-                lw a3,16(sp)
-                lw a4,20(sp)
-                lw a5,24(sp)
-                lw a6,28(sp)
-                lw a7,32(sp)
-                ",
-                $restore_fprs,
-                "
-                addi sp,sp,112
-                // 执行真正的函数
-                jr t1
-                ",
-                sym crate::relocation::dl_fixup,
-            )
-        }
-    };
+#[unsafe(naked)]
+pub(crate) extern "C" fn dl_runtime_resolve() {
+    core::arch::naked_asm!(
+        "
+        // 保存整数参数寄存器
+        // ra, a0-a7: 9 * 4 = 36 bytes
+        // 栈帧总大小设为 112 字节以保持 16 字节对齐
+        addi sp,sp,-112
+        sw ra,0(sp)
+        sw a0,4(sp)
+        sw a1,8(sp)
+        sw a2,12(sp)
+        sw a3,16(sp)
+        sw a4,20(sp)
+        sw a5,24(sp)
+        sw a6,28(sp)
+        sw a7,32(sp)
+        ",
+        #[cfg(target_feature = "d")]
+        "
+        fsd fa0,40(sp)
+        fsd fa1,48(sp)
+        fsd fa2,56(sp)
+        fsd fa3,64(sp)
+        fsd fa4,72(sp)
+        fsd fa5,80(sp)
+        fsd fa6,88(sp)
+        fsd fa7,96(sp)
+        ",
+        #[cfg(all(target_feature = "f", not(target_feature = "d")))]
+        "
+        fsw fa0,40(sp)
+        fsw fa1,44(sp)
+        fsw fa2,48(sp)
+        fsw fa3,52(sp)
+        fsw fa4,56(sp)
+        fsw fa5,60(sp)
+        fsw fa6,64(sp)
+        fsw fa7,68(sp)
+        ",
+        "
+        // 这两个是plt代码设置的
+        mv a0,t0
+        srli a1,t1,3
+        // 调用重定位函数
+        call {0}
+        // 恢复参数寄存器
+        mv t1,a0
+        lw ra,0(sp)
+        lw a0,4(sp)
+        lw a1,8(sp)
+        lw a2,12(sp)
+        lw a3,16(sp)
+        lw a4,20(sp)
+        lw a5,24(sp)
+        lw a6,28(sp)
+        lw a7,32(sp)
+        ",
+        #[cfg(target_feature = "d")]
+        "
+        fld fa0,40(sp)
+        fld fa1,48(sp)
+        fld fa2,56(sp)
+        fld fa3,64(sp)
+        fld fa4,72(sp)
+        fld fa5,80(sp)
+        fld fa6,88(sp)
+        fld fa7,96(sp)
+        ",
+        #[cfg(all(target_feature = "f", not(target_feature = "d")))]
+        "
+        flw fa0,40(sp)
+        flw fa1,44(sp)
+        flw fa2,48(sp)
+        flw fa3,52(sp)
+        flw fa4,56(sp)
+        flw fa5,60(sp)
+        flw fa6,64(sp)
+        flw fa7,68(sp)
+        ",
+        "
+        addi sp,sp,112
+        // 执行真正的函数
+        jr t1
+        ",
+        sym crate::relocation::dl_fixup,
+    )
 }
-
-#[cfg(target_feature = "d")]
-riscv32_dl_runtime_resolve!(
-    "
-    fsd fa0,40(sp)
-    fsd fa1,48(sp)
-    fsd fa2,56(sp)
-    fsd fa3,64(sp)
-    fsd fa4,72(sp)
-    fsd fa5,80(sp)
-    fsd fa6,88(sp)
-    fsd fa7,96(sp)
-    ",
-    "
-    fld fa0,40(sp)
-    fld fa1,48(sp)
-    fld fa2,56(sp)
-    fld fa3,64(sp)
-    fld fa4,72(sp)
-    fld fa5,80(sp)
-    fld fa6,88(sp)
-    fld fa7,96(sp)
-    "
-);
-
-#[cfg(all(target_feature = "f", not(target_feature = "d")))]
-riscv32_dl_runtime_resolve!(
-    "
-    fsw fa0,40(sp)
-    fsw fa1,44(sp)
-    fsw fa2,48(sp)
-    fsw fa3,52(sp)
-    fsw fa4,56(sp)
-    fsw fa5,60(sp)
-    fsw fa6,64(sp)
-    fsw fa7,68(sp)
-    ",
-    "
-    flw fa0,40(sp)
-    flw fa1,44(sp)
-    flw fa2,48(sp)
-    flw fa3,52(sp)
-    flw fa4,56(sp)
-    flw fa5,60(sp)
-    flw fa6,64(sp)
-    flw fa7,68(sp)
-    "
-);
-
-#[cfg(not(target_feature = "f"))]
-riscv32_dl_runtime_resolve!("", "");
 /// Static TLSDESC resolver dummy for RISC-V 32.
 pub(crate) extern "C" fn tlsdesc_resolver_static() {
     unimplemented!("TLSDESC is not supported on RISC-V 32 yet");