From d30a95f7d5d855ec1bb44dfec76ce4db1b494421 Mon Sep 17 00:00:00 2001
From: Nicholas Ward <npward@berkeley.edu>
Date: Thu, 1 Dec 2022 20:09:11 -0800
Subject: [PATCH] fixes

---
 .../cpu/kernel/asm/hash/blake/compression.asm |  9 ++-
 evm/src/cpu/kernel/asm/hash/blake/util.asm    | 72 ++++++++-----------
 2 files changed, 33 insertions(+), 48 deletions(-)

diff --git a/evm/src/cpu/kernel/asm/hash/blake/compression.asm b/evm/src/cpu/kernel/asm/hash/blake/compression.asm
index 01ac0aa5..eeb55458 100644
--- a/evm/src/cpu/kernel/asm/hash/blake/compression.asm
+++ b/evm/src/cpu/kernel/asm/hash/blake/compression.asm
@@ -3,7 +3,7 @@ global blake_compression:
     PUSH 0
     // stack: cur_block = 0, retdest
     %blake_initial_hash_value
-blake_compression_loop:
+compression_loop:
     // stack: h_0, ..., h_7, cur_block, retdest
     %blake_hash_value_addr
     // stack: addr, h_0, ..., h_7, cur_block, retdest
@@ -211,9 +211,9 @@ blake_compression_loop:
     // stack: num_blocks, cur_block + 1, h_0', h_1', h_2', h_3', h_4', h_5', h_6', h_7', cur_block + 1, retdest
     EQ
     // stack: last_block, h_0', h_1', h_2', h_3', h_4', h_5', h_6', h_7', cur_block + 1, retdest
-    %jumpi(blake_compression_end)
-    %jump(blake_compression_loop)
-blake_compression_end:
+    %jumpi(compression_end)
+    %jump(compression_loop)
+compression_end:
     // stack: h_0', h_1', h_2', h_3', h_4', h_5', h_6', h_7', cur_block + 1, retdest
     PUSH 0
     // stack: dummy=0, h_0', h_1', h_2', h_3', h_4', h_5', h_6', h_7', cur_block + 1, retdest
@@ -260,5 +260,4 @@ blake_compression_end:
     // stack: hash_second = h_4' || h_5' || h_6' || h_7', hash_first = h_0' || h_1' || h_2' || h_3', retdest
     %stack (second, first, ret) -> (ret, second, first)
     // stack: retdest, hash_first, hash_second
-    STOP
     JUMP
\ No newline at end of file
diff --git a/evm/src/cpu/kernel/asm/hash/blake/util.asm b/evm/src/cpu/kernel/asm/hash/blake/util.asm
index 2ce54672..49b78c1b 100644
--- a/evm/src/cpu/kernel/asm/hash/blake/util.asm
+++ b/evm/src/cpu/kernel/asm/hash/blake/util.asm
@@ -15,47 +15,33 @@
     // stack: (hi << 32) | lo
 %endmacro
 
-%macro invert_bytes_blake_word
-    // stack: word, ...
-    DUP1
-    %and_const(0xff)
-    %shl_const(56)
-    SWAP1
-    // stack: word, first_byte, ...
-    DUP1
-    %shr_const(8)
-    %and_const(0xff)
-    %shl_const(48)
-    SWAP1
-    // stack: word, second_byte, first_byte, ...
-    DUP1
-    %shr_const(16)
-    %and_const(0xff)
-    %shl_const(40)
-    SWAP1
-    DUP1
-    %shr_const(24)
-    %and_const(0xff)
-    %shl_const(32)
-    SWAP1
-    DUP1
-    %shr_const(32)
-    %and_const(0xff)
-    %shl_const(24)
-    SWAP1
-    DUP1
-    %shr_const(40)
-    %and_const(0xff)
-    %shl_const(16)
-    SWAP1
-    DUP1
-    %shr_const(48)
-    %and_const(0xff)
-    %shl_const(8)
-    SWAP1
-    %shr_const(56)
-    %and_const(0xff)
-    %rep 7
-        OR
-    %endrep
+// Invert the order of the four bytes in a word.
+%macro invert_four_byte_word
+    // stack: word
+    %mul_const(0x1000000010000000100)
+    %and_const(0xff0000ff00ff00000000ff0000)
+    %mod_const(0xffffffffffff)
+    // stack: word_inverted
+%endmacro
+
+// Invert the order of the eight bytes in a Blake word.
+%macro invert_bytes_blake_word
+    // stack: word
+    DUP1
+    // stack: word, word
+    %and_const(0xffffffff)
+    // stack: word_lo, word
+    SWAP1
+    // stack: word, word_lo
+    %shr_const(32)
+    // stack: word_hi, word_lo
+    %invert_four_byte_word
+    // stack: word_hi_inverted, word_lo
+    SWAP1
+    // stack: word_lo, word_hi_inverted
+    %invert_four_byte_word
+    // stack: word_lo_inverted, word_hi_inverted
+    %shl_const(32)
+    OR
+    // stack: word_inverted
 %endmacro