carry starting to make sense

diff --git a/openpower/sv/bitmanip/divmnu64.c b/openpower/sv/bitmanip/divmnu64.c
index 1249ddab5a995a8750ed9127d5320adc8fe5d826..7ab938ea78fb4bd47e6b646bb7de41df7eb49240 100644 (file)
--- a/openpower/sv/bitmanip/divmnu64.c
+++ b/openpower/sv/bitmanip/divmnu64.c
@@ -106,7 +106,7 @@ again:
         rhat = rhat + vn[n-1];
         if (rhat < b) goto again;
       }
-
+#define MUL_RSUB_CARRY_2_STAGE2
 #ifdef ORIGINAL
       // Multiply and subtract.
       k = 0;
@@ -220,6 +220,34 @@ again:
          un[i + j] = (uint32_t)result;
       }
       bool need_fixup = carry != 1;
+#elif defined(MUL_RSUB_CARRY_2_STAGE2)
+      (void)p; // shut up unused variable warning
+
+      // Multiply and subtract.
+      uint32_t carry = 0;
+      uint32_t phi[2000]; // plenty space
+      uint32_t plo[2000]; // plenty space
+      // same mul-and-sub as SUB_MUL_BORROW but not the same
+      // mul-and-sub-minus-one as MUL_RSUB_CARRY
+      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);
+         plo[i] = value & 0xffffffffLL;
+         phi[i] = value >> 32;
+      }
+      // NOW it starts to make sense. when no carry this time, next
+      // carry as-is. rlse next carry reduces by one.
+      // it here (as ~(0))
+      for(int i = 0; i <= n; i++) {
+         uint64_t result = (((uint64_t)phi[i]<<32) | plo[i]) + carry;
+         uint32_t result_high = result >> 32;
+         if(carry == 0)
+            carry = result_high;
+         else
+            carry = result_high-1;
+         un[i + j] = (uint32_t)result;
+      }
+      bool need_fixup = carry != 0;
 #else
 #error need to choose one of the algorithm options; e.g. -DORIGINAL
 #endif