Better instruction layout?

Author: EbbDrop

src/day25.rs

@@ -72,27 +72,27 @@ unsafe fn part1_inner(s: &[u8]) -> u64 {
                 "jb        6f",               // Jump to < 16 case
 
                 "4:",
+                "vpaddq    {vt}, {d},  ymmword ptr [{os} + 8*{i} - 32]",
                 "add       {i},  -16",
-                "vpaddq    {vt}, {d},  ymmword ptr [{os} + 8*{i} + 32*3]",
                 "vpand     {vt}, {vt}, {msb}",
                 "vpcmpeqq  {vt}, {vt}, {zero}",
                 "vpmovmskb {t},  {vt}",
-                "popcnt    {t},  {t}",
-                "add       {sum},{t}",
                 "vpaddq    {vt}, {d},  ymmword ptr [{os} + 8*{i} + 32*2]",
+                "popcnt    {t},  {t}",
                 "vpand     {vt}, {vt}, {msb}",
+                "add       {sum},{t}",
                 "vpcmpeqq  {vt}, {vt}, {zero}",
                 "vpmovmskb {t},  {vt}",
-                "popcnt    {t},  {t}",
-                "add       {sum},{t}",
                 "vpaddq    {vt}, {d},  ymmword ptr [{os} + 8*{i} + 32*1]",
+                "popcnt    {t},  {t}",
                 "vpand     {vt}, {vt}, {msb}",
+                "add       {sum},{t}",
                 "vpcmpeqq  {vt}, {vt}, {zero}",
                 "vpmovmskb {t},  {vt}",
-                "popcnt    {t},  {t}",
-                "add       {sum},{t}",
                 "vpaddq    {vt}, {d},  ymmword ptr [{os} + 8*{i} + 32*0]",
+                "popcnt    {t},  {t}",
                 "vpand     {vt}, {vt}, {msb}",
+                "add       {sum},{t}",
                 "vpcmpeqq  {vt}, {vt}, {zero}",
                 "vpmovmskb {t},  {vt}",
                 "popcnt    {t},  {t}",
@@ -105,8 +105,8 @@ unsafe fn part1_inner(s: &[u8]) -> u64 {
                 // Is >= 4 and < 16
 
                 "5:",
+                "vpaddq    {vt}, {d},  ymmword ptr [{os} + 8*{i} - 32]",
                 "add       {i}, -4",
-                "vpaddq    {vt}, {d},  ymmword ptr [{os} + 8*{i}]",
                 "vpand     {vt}, {vt}, {msb}",
                 "vpcmpeqq  {vt}, {vt}, {zero}",
                 "vpmovmskb {t},  {vt}",
@@ -143,34 +143,35 @@ unsafe fn part1_inner(s: &[u8]) -> u64 {
             keys_i += 1;
         } else {
             std::arch::asm!(
+
                 "test      {max_i}, {max_i}",
                 "je        2f",               // Jump on empty
                 "mov       {i}, {max_i}",
                 "cmp       {i}, 16",
                 "jb        6f",               // Jump to < 16 case
 
                 "4:",
+                "vpaddq    {vt}, {d},  ymmword ptr [{os} + 8*{i} - 32]",
                 "add       {i},  -16",
-                "vpaddq    {vt}, {d},  ymmword ptr [{os} + 8*{i} + 32*3]",
                 "vpand     {vt}, {vt}, {msb}",
                 "vpcmpeqq  {vt}, {vt}, {zero}",
                 "vpmovmskb {t},  {vt}",
-                "popcnt    {t},  {t}",
-                "add       {sum},{t}",
                 "vpaddq    {vt}, {d},  ymmword ptr [{os} + 8*{i} + 32*2]",
+                "popcnt    {t},  {t}",
                 "vpand     {vt}, {vt}, {msb}",
+                "add       {sum},{t}",
                 "vpcmpeqq  {vt}, {vt}, {zero}",
                 "vpmovmskb {t},  {vt}",
-                "popcnt    {t},  {t}",
-                "add       {sum},{t}",
                 "vpaddq    {vt}, {d},  ymmword ptr [{os} + 8*{i} + 32*1]",
+                "popcnt    {t},  {t}",
                 "vpand     {vt}, {vt}, {msb}",
+                "add       {sum},{t}",
                 "vpcmpeqq  {vt}, {vt}, {zero}",
                 "vpmovmskb {t},  {vt}",
-                "popcnt    {t},  {t}",
-                "add       {sum},{t}",
                 "vpaddq    {vt}, {d},  ymmword ptr [{os} + 8*{i} + 32*0]",
+                "popcnt    {t},  {t}",
                 "vpand     {vt}, {vt}, {msb}",
+                "add       {sum},{t}",
                 "vpcmpeqq  {vt}, {vt}, {zero}",
                 "vpmovmskb {t},  {vt}",
                 "popcnt    {t},  {t}",
@@ -183,8 +184,8 @@ unsafe fn part1_inner(s: &[u8]) -> u64 {
                 // Is >= 4 and < 16
 
                 "5:",
+                "vpaddq    {vt}, {d},  ymmword ptr [{os} + 8*{i} - 32]",
                 "add       {i}, -4",
-                "vpaddq    {vt}, {d},  ymmword ptr [{os} + 8*{i}]",
                 "vpand     {vt}, {vt}, {msb}",
                 "vpcmpeqq  {vt}, {vt}, {zero}",
                 "vpmovmskb {t},  {vt}",