Changeset 3911

Timestamp:

08/25/08 02:53:21 (3 months ago)

Author:

Don Clugston

Message:

Asm implementation of subMul. This speeds up division enormously.

Files:

• trunk/tango/math/internal/BignumX86.d (modified) (8 diffs)

trunk/tango/math/internal/BignumX86.d

@@ -426 +426 @@
-uint multibyteMulAdd(char op)(uint [] dest, uint[] src, uint multiplier, uint carry)
-{
-    static if (op=='-') {
-      /* This is equivalent to:
-        ---
-        uint [] tmp = new uint[src.length];
-        uint c = multibyteMul(tmp, src, multiplier, carry);
-        return c + multibyteAddSub!('-')(dest, dest, tmp, 0);
-        ---
-      */
-        ulong c = carry;
-        for (int i = 0; i < src.length; i++) {
-            c += cast(ulong)multiplier * src[i];
-            ulong t = cast(ulong)dest[i] - cast(uint)c;
-            dest[i] = cast(uint)t;
-            c = cast(uint)((c>>32) - (t>>32));        
-        }
-        return cast(uint)c;  
-   } else {
-
-
-    // Timing: This is the most time-critical bignum function.
-    // Pentium M: 5.4 cycles/operation, still has 2 resource stalls + 1 load block/iteration
@@ -467 +448 @@
-    // ESI = src
-    
+    const char [] OP = (op=='+')? "add" : "sub";
-    version(D_PIC) {
-        enum { zero = 0 }
@@ -478 +460 @@
-    
-    enum { LASTPARAM = 5*4 } // 4* pushes + return address.
+mixin("        
-    asm {
-        naked;
@@ -510 +493 @@
-
-        mul int ptr [ESP+LASTPARAM];
-
-
+
-        mov EBP, zero;
-    
@@ -523 +505 @@
-L1:
-        mul int ptr [ESP+LASTPARAM];
-add
+" ~ OP ~ "
-        mov ECX, zero;
- 
@@ -538 +520 @@
-    asm {        
-        mul int ptr [ESP+LASTPARAM];
-add
+" ~ OP ~ "
-        mov EBP, zero;
-    
@@ -547 +529 @@
-        add EBX, 2;
-        jl L1;
-
-
+
-        mov EAX, EBP; // get final carry
-        adc EAX, 0;
@@ -567 +548 @@
-        jl L1;
-        jmp L_done;
-
-
-
+
-}
-