Improve blake2f call (#1477)

* Improve on blake2 operations

* Comments

* Remove swap_mstore calls by changing stack macros

evm/src/cpu/kernel/asm/hash/blake2/addresses.asm

@@ -1,12 +1,20 @@
 // Address where the working version of the hash value is stored.
+// It is ready to be used, i.e. already containing the current context
+// and SEGMENT_KERNEL_GENERAL.
 %macro blake2_hash_value_addr
-    PUSH 0
-    // stack: 0
-    %mload_current_general
-    // stack: num_blocks
+    PUSH @SEGMENT_KERNEL_GENERAL
+    // stack: segment
+    GET_CONTEXT
+    // stack: context, segment
+    %build_address_no_offset
+    DUP1
+    MLOAD_GENERAL
+    // stack: num_blocks, addr
     %block_size
     %add_const(2)
-    // stack: num_bytes+2
+    // stack: num_bytes+2, addr
+    ADD
+    // stack: addr
 %endmacro
 
 // Address where the working version of the compression internal state is stored.

evm/src/cpu/kernel/asm/hash/blake2/blake2_f.asm

@@ -6,9 +6,9 @@ global blake2_f:
     // stack: addr, rounds, h0...h7, m0...m15, t0, t1, flag, retdest
     %rep 8
         // stack: addr, rounds, h_i, ...
-        %stack (addr, rounds, h_i) -> (addr, h_i, addr, rounds)
-        // stack: addr, h_i, addr, rounds, ...
-        %mstore_current_general
+        %stack (addr, rounds, h_i) -> (h_i, addr, addr, rounds)
+        // stack: h_i, addr, addr, rounds, ...
+        MSTORE_GENERAL
         %increment
     %endrep
 
@@ -21,9 +21,9 @@ global blake2_f:
     // stack: message_addr, rounds, m0...m15, t0, t1, flag, retdest
     %rep 16
         // stack: message_addr, rounds, m_i, ...
-        %stack (message_addr, rounds, m_i) -> (message_addr, m_i, message_addr, rounds)
-        // stack: message_addr, m_i, message_addr, rounds, ...
-        %mstore_current_general
+        %stack (message_addr, rounds, m_i) -> (m_i, message_addr, message_addr, rounds)
+        // stack: m_i, message_addr, message_addr, rounds, ...
+        MSTORE_GENERAL
         %increment
     %endrep
 
@@ -37,7 +37,7 @@ global blake2_f:
         // stack: addr, ...
         DUP1
         // stack: addr, addr, ...
-        %mload_current_general
+        MLOAD_GENERAL
         // stack: val, addr, ...
         SWAP1
         // stack: addr, val, ...
@@ -53,31 +53,30 @@ global blake2_f:
 
     // First eight words of the internal state: current hash value h_0, ..., h_7.
     %rep 8
-        SWAP1
-        DUP2
-        %mstore_current_general
+        DUP1
+        SWAP2
+        MSTORE_GENERAL
         %increment
     %endrep
     // stack: start + 8, rounds, t0, t1, flag, retdest
 
     // Next four values of the internal state: first four IV values.
     PUSH 0
-    // stack: 0, start + 8, rounds, t0, t1, flag, retdest
+    // stack: 0, addr, rounds, t0, t1, flag, retdest
     %rep 4
-        // stack: i, loc, ...
-        DUP1
-        // stack: i, i, loc, ...
+        // stack: i, addr, ...
+        DUP2
+        DUP2
+        // stack: i, addr, i, addr, ...
         %blake2_iv
-        // stack: IV_i, i, loc, ...
-        DUP3
-        // stack: loc, IV_i, i, loc, ...
-        %mstore_current_general
-        // stack: i, loc, ...
+        // stack: IV_i, addr, i, addr, ...
+        MSTORE_GENERAL
+        // stack: i, addr, ...
         %increment
         SWAP1
         %increment
         SWAP1
-        // stack: i + 1, loc + 1,...
+        // stack: i + 1, addr + 1,...
     %endrep
     // stack: 4, start + 12, rounds, t0, t1, flag, retdest
     POP
@@ -92,29 +91,28 @@ global blake2_f:
     // Last four values of the internal state: last four IV values, XOR'd with
     // the values (t0, t1, invert_if_flag, 0).
     %rep 4
-        // stack: i, loc, val, next_val,...
-        DUP1
-        // stack: i, i, loc, val, next_val,...
+        // stack: i, addr, val, next_val,...
+        DUP2
+        DUP2
+        // stack: i, addr, i, addr, val, next_val,...
         %blake2_iv
-        // stack: IV_i, i, loc, val, next_val,...
-        DUP4
-        // stack: val, IV_i, i, loc, val, next_val,...
+        // stack: IV_i, addr, i, addr, val, next_val,...
+        DUP5
+        // stack: val, IV_i, addr, i, addr, val, next_val,...
         XOR
-        // stack: val ^ IV_i, i, loc, val, next_val,...
-        DUP3
-        // stack: loc, val ^ IV_i, i, loc, val, next_val,...
-        %mstore_current_general
-        // stack: i, loc, val, next_val,...
+        // stack: val ^ IV_i, addr, i, addr, val, next_val,...
+        MSTORE_GENERAL
+        // stack: i, addr, val, next_val,...
         %increment
-        // stack: i + 1, loc, val, next_val,...
+        // stack: i + 1, addr, val, next_val,...
         SWAP2
-        // stack: val, loc, i + 1, next_val,...
+        // stack: val, addr, i + 1, next_val,...
         POP
-        // stack: loc, i + 1, next_val,...
+        // stack: addr, i + 1, next_val,...
         %increment
-        // stack: loc + 1, i + 1, next_val,...
+        // stack: addr + 1, i + 1, next_val,...
         SWAP1
-        // stack: i + 1, loc + 1, next_val,...
+        // stack: i + 1, addr + 1, next_val,...
     %endrep
     // stack: 8, start + 16, rounds, retdest
     %pop2

evm/src/cpu/kernel/asm/hash/blake2/g_functions.asm

@@ -11,28 +11,28 @@
     DUP11
     // stack: start, a, b, c, d, a, b, c, d, x, y, start
     ADD
-    %mload_current_general
+    MLOAD_GENERAL
     // stack: v[a], b, c, d, a, b, c, d, x, y, start
     SWAP1
     // stack: b, v[a], c, d, a, b, c, d, x, y, start
     DUP11
     // stack: start, b, v[a], c, d, a, b, c, d, x, y, start
     ADD
-    %mload_current_general
+    MLOAD_GENERAL
     // stack: v[b], v[a], c, d, a, b, c, d, x, y, start
     SWAP2
     // stack: c, v[a], v[b], d, a, b, c, d, x, y, start
     DUP11
     // stack: start, c, v[a], v[b], d, a, b, c, d, x, y, start
     ADD
-    %mload_current_general
+    MLOAD_GENERAL
     // stack: v[c], v[a], v[b], d, a, b, c, d, x, y, start
     SWAP3
     // stack: d, v[a], v[b], v[c], a, b, c, d, x, y, start
     DUP11
     // stack: start, d, v[a], v[b], v[c], a, b, c, d, x, y, start
     ADD
-    %mload_current_general
+    MLOAD_GENERAL
     // stack: v[d], v[a], v[b], v[c], a, b, c, d, x, y, start
     %stack (vd, vs: 3) -> (vs, vd)
     // stack: v[a], v[b], v[c], v[d], a, b, c, d, x, y, start
@@ -95,13 +95,13 @@
     %stack (vb, vc, vd, va, a, b, c, d, x, y, start) -> (start, a, va, start, b, vb, start, c, vc, start, d, vd)
     // stack: start, a, v[a]'', start, b, v[b]'', start, c, v[c]'', start, d, v[d]''
     ADD
-    %mstore_current_general
+    %swap_mstore
     ADD
-    %mstore_current_general
+    %swap_mstore
     ADD
-    %mstore_current_general
+    %swap_mstore
     ADD
-    %mstore_current_general
+    %swap_mstore
 %endmacro
 
 %macro call_blake2_g_function(a, b, c, d, x_idx, y_idx)
@@ -113,7 +113,7 @@
     // stack: s[y_idx], round, start
     %blake2_message_addr
     ADD
-    %mload_current_general
+    MLOAD_GENERAL
     // stack: m[s[y_idx]], round, start
     PUSH $x_idx
     DUP3
@@ -122,7 +122,7 @@
     // stack: s[x_idx], m[s[y_idx]], round, start
     %blake2_message_addr
     ADD
-    %mload_current_general
+    MLOAD_GENERAL
     // stack: m[s[x_idx]], m[s[y_idx]], round, start
     %stack (ss: 2, r, s) -> (ss, s, r, s)
     // stack: m[s[x_idx]], m[s[y_idx]], start, round, start

evm/src/cpu/kernel/asm/hash/blake2/hash.asm

@@ -5,13 +5,13 @@ blake2_generate_new_hash_value:
     // stack: addr, i, retdest
     DUP2
     ADD
-    %mload_current_general
+    MLOAD_GENERAL
     // stack: h_i, i, retdest
     %blake2_internal_state_addr
     // stack: addr, h_i, i, retdest
     DUP3
     ADD
-    %mload_current_general
+    MLOAD_GENERAL
     // stack: v_i, h_i, i, retdest
     %blake2_internal_state_addr
     // stack: addr, v_i, h_i, i, retdest
@@ -21,7 +21,7 @@ blake2_generate_new_hash_value:
     // stack: i, addr, h_i, v_i, retdest
     ADD
     %add_const(8)
-    %mload_current_general
+    MLOAD_GENERAL
     // stack: v_(i+8), h_i, v_i, retdest
     XOR
     XOR