Fixing subnormal result discarded bit masking

Signed-off-by: Nicolas Brunie <nibrunie@gmail.com>

Author: nibrunie

src/unpriv/rvbna.adoc

@@ -231,9 +231,10 @@ BulkNormalizedDotProduct(A[n], B[n]) {
         } else {
             // denormalization and final round-to-odd
             // (of bits discarded during denormalization)
-            let denormalizedSig = accAbs >> (q - 1 + resExp)
-            let discardedMask = ((1 << (q - 1)) - 1) >> (q - 1 + resExp)
-            let discardedBits = accAbs & discardedMask
+            let denormShift = -resExp
+            let denormalizedSig = (accAbs << lzc) >> (g + o + 1 + 1 + denormShift)
+            let discardedMask = (1 << (g + o + 1 + 1 + denormShift)) - 1
+            let discardedBits = (accAbs << lzc) & discardedMask
             let forceLSB =  (discardedBits != 0 ? 1 : 0)
             return (accSign << (q + f - 1)) | denormalizedSig | forceLSB
         }