diff --git a/arch/aarch64/defs.h b/arch/aarch64/defs.h
index ee9fc7f..1282f56 100644
--- a/arch/aarch64/defs.h
+++ b/arch/aarch64/defs.h
@@ -9,6 +9,7 @@
 #define SP_OFFSET	432
 #define PC_OFFSET	440
 #define PSTATE_OFFSET	448
+#define D8_OFFSET 456
 
 #ifndef FPSIMD_MAGIC
 # define FPSIMD_MAGIC	0x46508001
diff --git a/arch/aarch64/getcontext.S b/arch/aarch64/getcontext.S
index 1098706..a4a8bfa 100644
--- a/arch/aarch64/getcontext.S
+++ b/arch/aarch64/getcontext.S
@@ -50,7 +50,10 @@ PROC_NAME(libucontext_getcontext):
 	/* save pstate */
 	str	xzr, [x0, #PSTATE_OFFSET]
 
-	/* TODO: SIMD / FPRs */
+	stp d8, d9,   [x0, #D8_OFFSET +  0]
+	stp d10, d11, [x0, #D8_OFFSET + 16]
+	stp d12, d13, [x0, #D8_OFFSET + 32]
+	stp d14, d15, [x0, #D8_OFFSET + 48]
 
 	mov	x0, #0
 	ret
diff --git a/arch/aarch64/include/libucontext/bits.h b/arch/aarch64/include/libucontext/bits.h
index e249448..0bc0563 100644
--- a/arch/aarch64/include/libucontext/bits.h
+++ b/arch/aarch64/include/libucontext/bits.h
@@ -14,7 +14,8 @@ typedef struct sigcontext {
 	unsigned long fault_address;
 	unsigned long regs[31];
 	unsigned long sp, pc, pstate;
-	long double __reserved[256];
+	unsigned long d8_through_d15[8];
+	long double __reserved[252];
 } libucontext_mcontext_t;
 
 typedef struct {
diff --git a/arch/aarch64/setcontext.S b/arch/aarch64/setcontext.S
index 3d455a5..cbe1ddc 100644
--- a/arch/aarch64/setcontext.S
+++ b/arch/aarch64/setcontext.S
@@ -33,7 +33,10 @@ PROC_NAME(libucontext_setcontext):
 	ldr	x2, [x0, #SP_OFFSET]
 	mov	sp, x2
 
-	/* TODO: SIMD / FPRs */
+	ldp d8, d9,   [x0, #D8_OFFSET +  0]
+	ldp d10, d11, [x0, #D8_OFFSET + 16]
+	ldp d12, d13, [x0, #D8_OFFSET + 32]
+	ldp d14, d15, [x0, #D8_OFFSET + 48]
 
 	/* save current program counter in link register */
 	ldr	x16, [x0, #PC_OFFSET]
diff --git a/arch/aarch64/swapcontext.S b/arch/aarch64/swapcontext.S
index fddfa41..e90f7f1 100644
--- a/arch/aarch64/swapcontext.S
+++ b/arch/aarch64/swapcontext.S
@@ -49,6 +49,11 @@ PROC_NAME(libucontext_swapcontext):
 	/* save pstate */
 	str	xzr, [x0, #PSTATE_OFFSET]
 
+	stp d8, d9,   [x0, #D8_OFFSET +  0]
+	stp d10, d11, [x0, #D8_OFFSET + 16]
+	stp d12, d13, [x0, #D8_OFFSET + 32]
+	stp d14, d15, [x0, #D8_OFFSET + 48]
+
 	/* context to swap to is in x1 so... we move to x0 and call setcontext */
 	/* store our link register in x28 */
 	mov	x28, x30