diff --git a/bin/chainboot/src/boot.rs b/bin/chainboot/src/boot.rs
index c1c44be0..8485e7f4 100644
--- a/bin/chainboot/src/boot.rs
+++ b/bin/chainboot/src/boot.rs
@@ -28,81 +28,8 @@ pub unsafe extern "C" fn _start() -> ! {
     unsafe { reset() };
 }
 
-#[unsafe(no_mangle)]
-#[unsafe(link_section = ".text.chainboot")]
-pub unsafe extern "C" fn reset() -> ! {
-    use core::{
-        cell::UnsafeCell,
-        sync::{atomic, atomic::Ordering},
-    };
-
-    // These are a problem, because they are not interpreted as constants here.
-    // Subsequently, this code tries to read values from not-yet-existing data locations.
-    unsafe extern "Rust" {
-        // Boundaries of the .bss section, provided by the linker script
-        static __BSS_START: UnsafeCell<()>;
-        static __BSS_SIZE_U64S: UnsafeCell<()>;
-        // Load address of the kernel binary
-        static __binary_nonzero_lma: UnsafeCell<()>;
-        // Address to relocate to and image size
-        static __binary_nonzero_vma: UnsafeCell<()>;
-        static __binary_nonzero_vma_end_exclusive: UnsafeCell<()>;
-        // Stack top
-        static __boot_core_stack_end_exclusive: UnsafeCell<()>;
-    }
-
-    // This tries to call memcpy() at a wrong linked address - the function is in relocated area!
-
-    // Relocate the code.
-    // Emulate
-    // core::ptr::copy_nonoverlapping(
-    //     __binary_nonzero_lma.get() as *const u64,
-    //     __binary_nonzero_vma.get() as *mut u64,
-    //     __binary_nonzero_vma_end_exclusive.get() as usize - __binary_nonzero_vma.get() as usize,
-    // );
-    let binary_size = unsafe { __binary_nonzero_vma_end_exclusive.get() } as usize
-        - unsafe { __binary_nonzero_vma.get() } as usize;
-    unsafe {
-        local_memcpy(
-            __binary_nonzero_vma.get() as *mut u8,
-            __binary_nonzero_lma.get() as *const u8,
-            binary_size,
-        )
-    };
-
-    // This tries to call memset() at a wrong linked address - the function is in relocated area!
-
-    // Zeroes the .bss section
-    // Emulate
-    // crate::stdmem::local_memset(__bss_start.get() as *mut u8, 0u8, __bss_size.get() as usize);
-    let bss = unsafe {
-        core::slice::from_raw_parts_mut(
-            __BSS_START.get() as *mut u64,
-            __BSS_SIZE_U64S.get() as usize,
-        )
-    };
-    for i in bss {
-        *i = 0;
-    }
-
-    // Don't cross this line with loads and stores. The initializations
-    // done above could be "invisible" to the compiler, because we write to the
-    // same memory location that is used by statics after this point.
-    // Additionally, we assume that no statics are accessed before this point.
-    atomic::compiler_fence(Ordering::SeqCst);
-
-    let max_kernel_size = unsafe { __binary_nonzero_vma.get() } as u64
-        - unsafe { __boot_core_stack_end_exclusive.get() } as u64;
-    unsafe { crate::kernel_init(max_kernel_size) }
+unsafe extern "Rust" {
+    fn reset() -> !;
 }
 
-#[inline(always)]
-#[unsafe(link_section = ".text.chainboot")]
-unsafe fn local_memcpy(mut dest: *mut u8, mut src: *const u8, n: usize) {
-    let dest_end = unsafe { dest.add(n) };
-    while dest < dest_end {
-        unsafe { *dest = *src };
-        dest = unsafe { dest.add(1) };
-        src = unsafe { src.add(1) };
-    }
-}
+core::arch::global_asm!(include_str!("boot.s"));
diff --git a/bin/chainboot/src/boot.s b/bin/chainboot/src/boot.s
new file mode 100644
index 00000000..06d77beb
--- /dev/null
+++ b/bin/chainboot/src/boot.s
@@ -0,0 +1,68 @@
+/*
+ * Pre-boot code.
+ * Used only because Rust's AM considers UB any access to statics before statics
+ * have been initialized. This is exactly the case for the boot code.
+ * So we avoid referencing any statics in the Rust code, and delegate the
+ * task to assembly piece instead.
+ */
+
+ // Load the address of a symbol into a register, PC-relative.
+ //
+ // The symbol must lie within +/- 4 GiB of the Program Counter.
+ //
+ // # Resources
+ //
+ // - https://sourceware.org/binutils/docs-2.36/as/AArch64_002dRelocations.html
+ .macro ADR_REL register, symbol
+	adrp \register, \symbol
+	add  \register, \register, #:lo12:\symbol
+ .endm
+
+// .section .text.chainboot.entry
+
+.section .text.chainboot
+/// Reset function.
+///
+/// Initializes the bss section before calling into the user's `main()`.
+///
+/// # Safety
+///
+/// We assume that no statics are accessed before transition to main from this function.
+///
+/// We are guaranteed to be in EL1 non-secure mode here.
+reset:
+    ADR_REL x0, __boot_core_stack_end_exclusive
+    mov sp, x0
+
+    // Relocate the code from __binary_nonzero_lma to __binary_nonzero_vma
+    ADR_REL x1, __binary_nonzero_lma      // Load address of the kernel binary
+    ADR_REL x2, __binary_nonzero_vma      // Address to relocate to
+    ADR_REL x3, __binary_nonzero_vma_end_exclusive // To calculate image size
+
+    sub x0, x2, x0 // max loadable kernel size = VMA - SP
+
+    // Relocate the code.
+    sub x4, x3, x2 // x4 = Image size
+
+.L__relocate_loop:
+    ldp x5, x6, [x1], #16
+    stp x5, x6, [x2], #16
+    sub x4, x4, #16
+    b.nz .L__relocate_loop
+
+    // Initialize BSS
+    // Assumptions: BSS start is u64-aligned, BSS end is u128-aligned.
+    // __BSS_START and __BSS_END are defined in linker script
+    ADR_REL x1, __BSS_START
+    ADR_REL x2, __BSS_END
+.L__bss_init_loop:
+    stp xzr, xzr, [x1], #16
+    cmp x1, x2
+    b.lt .L__bss_init_loop
+
+    // max_kernel_size is already in x0 here
+    b kernel_init
+
+.size	reset, . - reset
+.type	reset, function
+.global	reset
diff --git a/bin/chainboot/src/link.ld b/bin/chainboot/src/link.ld
index 84748881..cff3505e 100644
--- a/bin/chainboot/src/link.ld
+++ b/bin/chainboot/src/link.ld
@@ -55,8 +55,8 @@ SECTIONS
         *(.text.chainboot)
     } :segment_start_code
 
-    /* Align to 8 bytes, b/c relocating the binary is done in u64 chunks */
-    . = ALIGN(8);
+    /* Align to 16 bytes, b/c relocating the binary is done in u128 chunks */
+    . = ALIGN(16);
 
     __binary_nonzero_lma = .;
 
@@ -88,8 +88,8 @@ SECTIONS
     ***********************************************************************************************/
     .data : { *(.data*) } :segment_data
 
-    /* Fill up to 8 bytes, b/c relocating the binary is done in u64 chunks */
-    . = ALIGN(8);
+    /* Fill up to 16 bytes, b/c relocating the binary is done in u128 chunks */
+    . = ALIGN(16);
     __binary_nonzero_vma_end_exclusive = .;
 
     /* Section is zeroed in pairs of u64. Align start and end to 16 bytes at least */
@@ -99,7 +99,7 @@ SECTIONS
         *(.bss .bss.*)
         *(COMMON)
         . = ALIGN(16);
-        __BSS_SIZE_U64S = (. - __BSS_START) / 8;
+        __BSS_END = .;
     } :segment_data
 
     /DISCARD/ : { *(.comment) *(.gnu*) *(.note*) *(.eh_frame*) *(.text.boot*)}
diff --git a/bin/chainboot/src/main.rs b/bin/chainboot/src/main.rs
index 3af39cf8..3375d7e5 100644
--- a/bin/chainboot/src/main.rs
+++ b/bin/chainboot/src/main.rs
@@ -23,6 +23,7 @@ mod boot;
 ///
 /// - Only a single core must be active and running this function.
 /// - The init calls in this function must appear in the correct order.
+#[unsafe(no_mangle)]
 unsafe fn kernel_init(max_kernel_size: u64) -> ! {
     #[cfg(feature = "jtag")]
     machine::debug::jtag::wait_debugger();