https://bugs.libre-soc.org/show_bug.cgi?id=817#c31

use 2-step 64/32 divides (experiment)

diff --git a/openpower/sv/biginteger/divmnu64.c b/openpower/sv/biginteger/divmnu64.c
index c77d9b780a929aba11d7a39107cc5294876709a1..a243bcd96349721b198e2021ebe594296293d257 100644 (file)
--- a/openpower/sv/biginteger/divmnu64.c
+++ b/openpower/sv/biginteger/divmnu64.c
@@ -125,10 +125,15 @@ int divmnu(unsigned q[], unsigned r[], const unsigned u[], const unsigned v[],
 
     for (j = m - n; j >= 0; j--)
     { // Main loop.
-        // Compute estimate qhat of q[j] from top 2 digits
-        uint64_t dig2 = ((uint64_t)un[j + n] << 32) | un[j + n - 1];
-        qhat = dig2 / vn[n - 1];
-        rhat = dig2 % vn[n - 1];
+        // Compute estimate qhat of q[j] from top 2 digits.
+        // do as 2 separate divs to demo that 64/32 div/rem would
+        // be perfectly fine.
+        uint64_t dig1 = (uint64_t)un[j + n];
+        qhat = dig1 / vn[n - 1];
+        rhat = dig1 % vn[n - 1];
+        uint64_t dig2 = ((uint64_t)rhat << 32) | un[j + n - 1];
+        qhat = dig2 / vn[n - 1] | (qhat<<32);
+        rhat = dig2 % vn[n - 1] | (rhat<<32);
     again:
         // use 3rd-from-top digit to obtain better accuracy
         if (qhat >= b || qhat * vn[n - 2] > b * rhat + un[j + n - 2])