add requested loop variants

diff --git a/openpower/sv/bitmanip/divmnu64.c b/openpower/sv/bitmanip/divmnu64.c
index 1d3585eec60139aaeb377044887aa15cdff459b0..93735d8027fdcbdb22d9d759eb6e61b66cd24f53 100644 (file)
--- a/openpower/sv/bitmanip/divmnu64.c
+++ b/openpower/sv/bitmanip/divmnu64.c
@@ -9,6 +9,7 @@ machine which has a 64/64 division instruction. */
 #include <stdio.h>
 #include <stdlib.h>     //To define "exit", req'd by XLC.
 #include <stdbool.h>
+#include <stdint.h>
 
 #define max(x, y) ((x) > (y) ? (x) : (y))
 
@@ -106,6 +107,7 @@ again:
         if (rhat < b) goto again;
       }
 
+#ifdef ORIGINAL
       // Multiply and subtract.
       k = 0;
       for (i = 0; i < n; i++) {
@@ -117,6 +119,36 @@ again:
       t = un[j+n] - k;
       un[j+n] = t;
       bool need_fixup = t < 0;
+#elif defined(SUB_MUL_BORROW)
+      (void)p; // shut up unused variable warning
+
+      // Multiply and subtract.
+      uint32_t borrow = 0;
+      for(int i = 0; i <= n; i++) {
+         uint32_t vn_i = i < n ? vn[i] : 0;
+         uint64_t value = un[i + j] - (uint64_t)qhat * vn_i - borrow;
+         borrow = -(uint32_t)(value >> 32);
+         un[i + j] = (uint32_t)value;
+      }
+      bool need_fixup = borrow != 0;
+#elif defined(MUL_RSUB_CARRY)
+      (void)p; // shut up unused variable warning
+
+      // Multiply and subtract.
+      uint32_t carry = 1;
+      for(int i = 0; i <= n; i++) {
+         uint32_t vn_i = i < n ? vn[i] : 0;
+         uint64_t result = un[i + j] + ~((uint64_t)qhat * vn_i) + carry;
+         uint32_t result_high = result >> 32;
+         if(carry <= 1)
+            result_high++;
+         carry = result_high;
+         un[i + j] = (uint32_t)result;
+      }
+      bool need_fixup = carry != 1;
+#else
+#error need to define one of ORIGINAL, SUB_MUL_BORROW, or MUL_RSUB_CARRY
+#endif
 
       q[j] = qhat;              // Store quotient digit.
       if (need_fixup) {         // If we subtracted too