diff --git a/src/jit_compiler_a64_static.S b/src/jit_compiler_a64_static.S
index bad9db88..ce1a690f 100644
--- a/src/jit_compiler_a64_static.S
+++ b/src/jit_compiler_a64_static.S
@@ -160,8 +160,8 @@ DECL(randomx_program_aarch64):
 	mrs	x8, fpcr
 	rbit	x8, x8
 
-	# Save x0
-	str	x0, [sp, -16]!
+	# Save x0 and reserve stack space for light mode
+	str	x0, [sp, -112]!
 
 	# Read literals
 	ldr	x0, literal_x0
@@ -366,7 +366,7 @@ DECL(randomx_program_aarch64_update_spMix1):
 	bne	DECL(randomx_program_aarch64_main_loop)
 	
 	# Restore x0
-	ldr	x0, [sp], 16
+	ldr	x0, [sp], 112
 
 	# Store integer registers
 	stp	x4, x5, [x0, 0]
@@ -398,9 +398,8 @@ DECL(randomx_program_aarch64_update_spMix1):
 	ret
 
 DECL(randomx_program_aarch64_vm_instructions_end_light):
-	sub	sp, sp, 96
-	stp	x0, x1, [sp, 64]
-	stp	x2, x30, [sp, 80]
+	stp	x0, x1, [sp, 80]
+	stp	x2, x30, [sp, 96]
 
 	# mx ^= r[readReg2] ^ r[readReg3];
 	eor	x9, x9, x20
@@ -412,7 +411,7 @@ DECL(randomx_program_aarch64_vm_instructions_end_light):
 	mov	x0, x1
 
 	# x1 -> pointer to output
-	mov	x1, sp
+	add	x1, sp, 16
 
 DECL(randomx_program_aarch64_light_cacheline_align_mask):
 	# Actual mask will be inserted by JIT compiler
@@ -428,10 +427,9 @@ DECL(randomx_program_aarch64_light_dataset_offset):
 
 	bl	rx_calc_dataset_item
 
-	mov	x10, sp
-	ldp	x0, x1, [sp, 64]
-	ldp	x2, x30, [sp, 80]
-	add	sp, sp, 96
+	add	x10, sp, 16
+	ldp	x0, x1, [sp, 80]
+	ldp	x2, x30, [sp, 96]
 
 	b	rx_program_xor_with_dataset_line