From 4d1a0ca8c06ee558a92c9981a572bb962f73a6f6 Mon Sep 17 00:00:00 2001
From: dam <dam@gudinoff>
Date: Sun, 2 Jul 2023 20:02:26 +0100
Subject: Tweaked signed add.

---
 Math_Ext.jai | 47 +++++++++++++++++++++++++++++++++--------------
 1 file changed, 33 insertions(+), 14 deletions(-)

(limited to 'Math_Ext.jai')

diff --git a/Math_Ext.jai b/Math_Ext.jai
index 500962f..368779f 100644
--- a/Math_Ext.jai
+++ b/Math_Ext.jai
@@ -74,35 +74,53 @@ add :: (x: $Tx, y: $Ty, $USE_GENERIC: bool = false) -> result: $Tr, saturated: b
         result: Tr = ---;
         saturated: bool = ---;
 
+
         S_ADD_ASM :: #string DONE
         #asm {
-            // Calculate limit based on x's sign.
-            mov         limit: gpr, MAX;
-            mov         sign: gpr, x;
-            shr.SIZE    sign, BITS;
-            add.SIZE    limit, sign;            // If sign is 1, then limit will overflow from MAX to MIN.
             
-            mov         result, x;
-            add.SIZE    result, y;
-            seto        saturated;
-            cmovo       result, limit;
+            // Performance
+            // s8    | s16   | s32   | s64
+            // 1.243 | 1.242 | 1.215 | 1.210
+            //
+            mov         result, -1;             // Pre-set result with signed maximum: set all ones.
+            shr.SIZE    result, 1;              // Pre-set result with signed maximum: insert zero on MSB.
+            bt          x, SHIFT;               // Test signal bit (affect CF).
+            adc         result, 0;              // Overflow signed maximum to signed minimum if CF is set.
+            
+            add.SIZE    x, y;                   // Add values (affect OF).
+            seto        saturated;              // Set saturated flat if OF.
+            cmovno      result, x;              // Move add-result to result if NOT OF.
+            
+            // Performance - a bit of improvement... with some more code.
+            // s8    | s16   | s32   | s64
+            // 1.336 | 1.305 | 1.217 | 1.210
+            //
+            // mov         sign: gpr, x;               // Copy x value to sign variable.
+            // mov         limit: gpr, MAX;            // Pre-set limit with signed maximum.
+            // shr.SIZE    sign, SHIFT;                // Get sign of x value.
+            // add.SIZE    limit, sign;                // If sign is 1, overflow from signed maximum to signed minimum.
+            // 
+            // mov         result, x;                  // Copy x value to result.
+            // add.SIZE    result, y;                  // Add values (affect OF).
+            // seto        saturated;                  // Set saturated flag if OF.
+            // cmovo       result, limit;              // Move limit to result if OF.
         }
         DONE
 
         #if Tr == s8
-            #insert #run replace(replace(replace(S_ADD_ASM, ".SIZE", ".b"), "MAX", "127"), "BITS", "7");
+            #insert #run replace(replace(S_ADD_ASM, ".SIZE", ".b"), "SHIFT", "7"); // , "MAX", "127")
         #if Tr == s16
-            #insert #run replace(replace(replace(S_ADD_ASM, ".SIZE", ".w"), "MAX", "32767"), "BITS", "15");
+            #insert #run replace(replace(S_ADD_ASM, ".SIZE", ".w"), "SHIFT", "15"); // , "MAX", "32767")
         #if Tr == s32
-            #insert #run replace(replace(replace(S_ADD_ASM, ".SIZE", ".d"), "MAX", "2147483647"), "BITS", "31");
+            #insert #run replace(replace(S_ADD_ASM, ".SIZE", ".d"), "SHIFT", "31"); // , "MAX", "2147483647")
         #if Tr == s64
-            #insert #run replace(replace(replace(S_ADD_ASM, ".SIZE", ".q"), "MAX", "9223372036854775807"), "BITS", "63");
+            #insert #run replace(replace(S_ADD_ASM, ".SIZE", ".q"), "SHIFT", "63"); // , "MAX", "9223372036854775807")
 
         
         U_ADD_ASM :: #string DONE
         #asm {
-            add.SIZE    x, y;                   // Add values.
             mov         result, -1;             // Pre-set result with unsigned maximum.
+            add.SIZE    x, y;                   // Add values (affect CF).
             setc        saturated;              // Set saturated flag if CF.
             cmovnc      result, x;              // Move add-result to result if NOT CF.
         }
@@ -152,6 +170,7 @@ sub :: (x: $Tx, y: $Ty, $USE_GENERIC: bool = false) -> result: $Tr, overflow: bo
         result: Tr = ---;
         saturated: bool = ---;
 
+
         S_SUB_ASM :: #string DONE
         #asm {
             // Calculate limit based on x's sign.
-- 
cgit v1.2.3