diff --git a/evm/src/cpu/kernel/aggregator.rs b/evm/src/cpu/kernel/aggregator.rs
index 85ebd11a..d8466894 100644
--- a/evm/src/cpu/kernel/aggregator.rs
+++ b/evm/src/cpu/kernel/aggregator.rs
@@ -48,6 +48,10 @@ pub(crate) fn combined_kernel() -> Kernel {
         include_str!("asm/secp256k1/moddiv.asm"),
         include_str!("asm/secp256k1/lift_x.asm"),
         include_str!("asm/secp256k1/inverse_scalar.asm"),
+        include_str!("asm/sha2.asm"),
+        include_str!("asm/sha2_constants.asm"),
+        include_str!("asm/sha2_memory.asm"),
+        include_str!("asm/sha2_ops.asm"),
         include_str!("asm/ecrecover.asm"),
         include_str!("asm/rlp/encode.asm"),
         include_str!("asm/rlp/decode.asm"),
diff --git a/evm/src/cpu/kernel/asm/sha2.asm b/evm/src/cpu/kernel/asm/sha2.asm
index 37594f8d..10a2d815 100644
--- a/evm/src/cpu/kernel/asm/sha2.asm
+++ b/evm/src/cpu/kernel/asm/sha2.asm
@@ -123,7 +123,7 @@ global sha2_pad:
     // stack: num_bytes, retdest
     // STEP 3: insert num_blocks at start
     push 64
-    swap
+    swap1
     div
     %increment
     // stack: num_blocks := num_bytes // 64 + 1, retdest
diff --git a/evm/src/cpu/kernel/asm/sha2_ops.asm b/evm/src/cpu/kernel/asm/sha2_ops.asm
index cfb24d6a..c600afe9 100644
--- a/evm/src/cpu/kernel/asm/sha2_ops.asm
+++ b/evm/src/cpu/kernel/asm/sha2_ops.asm
@@ -54,7 +54,7 @@
     add
 %endmacro
 
-global sha2_sigma_0:
+%macro sha2_sigma_0
     JUMPDEST
     // stack: x
     dup1
@@ -76,6 +76,7 @@ global sha2_sigma_0:
     // stack: shr(x, 3), rotr(x, 18), rotr(x, 7)
     xor
     xor
+%endmacro
 
 %macro sha2_sigma_1
     // stack: x