Using generic ASM code for signed integer add.

1 files changed, 68 insertions, 8 deletions

diff --git a/Math_Ext.jai b/Math_Ext.jai
index bf1e8d6..44c4d44 100644
--- a/Math_Ext.jai
+++ b/Math_Ext.jai
@@ -34,7 +34,10 @@ main :: () {
     // add_test(cast(u8)1,       cast(u8)2,   3, u8, false);
     // add_test(cast(u8)255,       cast(u8)1,   255, u8, true);
 
-    add_test(cast(s8)S8_MAX,       cast(s8)1,   S8_MAX, s8, true);
+    add_test(cast( s8) S8_MAX,  cast( s8)1,  S8_MAX,     s8, true);
+    add_test(cast(s16)S16_MAX,  cast(s16)1, S16_MAX,    s16, true);
+    add_test(cast(s32)S32_MAX,  cast(s32)1, S32_MAX,    s32, true);
+    add_test(cast(s64)S64_MAX,  cast(s64)1, S64_MAX,    s64, true);
     // add_test(cast(s32)66,       cast(s64)-2,   64, s64, false);
     // add_test(cast(u32)66,       cast(s64)4,    70, s64, false);
     // add_test(cast(s32)S32_MAX,  cast(s64)1,    2147483648, s64, false);
@@ -51,7 +54,7 @@ main :: () {
 
     return;
     
-
+/*
     best_generic: float;
     best_asm: float;
     for 0..100 {
@@ -88,6 +91,7 @@ main :: () {
 
         return SUM_SIZE, time, time_asm;
     }
+    */
 
 }
 
@@ -140,7 +144,7 @@ old_add :: (x: $Tx, y: $Ty) -> result: $Tr, overflow: bool // #dump
     return x + y, false;
 }
 
-add :: (x: $Tx, y: $Ty) -> result: $Tr, overflow: bool // #dump
+add :: (x: $Tx, y: $Ty) -> result: $Tr, overflow: bool #dump
     #modify {
         type_info_x := cast(*Type_Info)Tx;
         type_info_y := cast(*Type_Info)Ty;
@@ -193,20 +197,75 @@ add :: (x: $Tx, y: $Ty) -> result: $Tr, overflow: bool // #dump
         result: Tr = ---;
         overflow: bool = ---;
 
-        #if Tr == s8 {
+#import "String";
+
+    ADD_ASM :: #string DONE
         #asm { // s8
             x === a;
             y === b;
-            mov     d: gpr === d, MAX;
-            mov.b     c: gpr === c, x; // TODO Not using .b was erroing.
+            mov     max: gpr === d, MAX;
+            mov.SIZE     c: gpr === c, x; // TODO Not using .b was erroing.
             shr     c, BITS;
-            add     c, d;
-            add.b   x, y; // add.b for 8bits
+            add     c, max;
+            add.SIZE   x, y; // add.b for 8bits
             cmovo   x, c;
             seto    overflow;
             mov     result, x;
         }
+    DONE
+
+    #if Tr == s8
+        #insert #run replace(replace(replace(ADD_ASM, ".SIZE", ".b"), "MAX", "127"), "BITS", "7");
+    #if Tr == s16
+        #insert #run replace(replace(replace(ADD_ASM, ".SIZE", ".w"), "MAX", "32767"), "BITS", "15");
+    #if Tr == s32
+        #insert #run replace(replace(replace(ADD_ASM, ".SIZE", ".d"), "MAX", "2147483647"), "BITS", "31");
+    #if Tr == s64
+        #insert #run replace(replace(replace(ADD_ASM, ".SIZE", ".q"), "MAX", "9223372036854775807"), "BITS", "63");
+
+
+    // TODO
+    // WIP - doing a similar process for unsigned saturation add
+        else #if Tr == u8 {
+        #asm { // u8
+            x === a;
+            y === b;
+            d: gpr;
+            add.b   x, y; // add.b for 8bits
+            setc    overflow;
+            sbb     d, d;
+            or      d, x;
+            mov     result, d;
+        }
         }
+        else #if Tr == u16 {
+        #asm { // u16
+            x === a;
+            y === b;
+            d: gpr;
+            add.w   x, y; // add.w for 16bits
+            setc    overflow;
+            sbb     d, d;
+            or      d, x;
+            mov     result, d;
+        }
+        }
+
+
+        // #asm { // s8
+            // x === a;
+            // y === b;
+            // mov     d: gpr === d, MAX;
+            // mov.b     c: gpr === c, x; // TODO Not using .b was erroing.
+            // shr     c, BITS;
+            // add     c, d;
+            // add.b   x, y; // add.b for 8bits
+            // cmovo   x, c;
+            // seto    overflow;
+            // mov     result, x;
+        // }
+        // }
+        /*
         else #if Tr == s32 {
         #asm { // s32
             x === a;
@@ -263,6 +322,7 @@ add :: (x: $Tx, y: $Ty) -> result: $Tr, overflow: bool // #dump
             mov     result, d;
         }
         }
+        */
         return result, overflow;
     }
 }