rvv: fix INT_ROUNDING compliance

* round-to-nearest-even: In the case that result[gb] = 0, the result
  should still be rounded up if result[gb-1] != 0 && result[gb-2:0] != 0
  (the usual round-to-nearest behavior outside of the tiebreaker).
* round-down: Since all uses of INT_ROUNDING() are immediately followed
  with a right shift by gb, clearing the lower bits is unnecessary.
* round-to-odd: The LSB should be OR'd only if result[gb-1:0] != 0.

Signed-off-by: Albert Ou <aou@eecs.berkeley.edu>

riscv/decode.h

@@ -476,31 +476,27 @@ static inline bool is_overlaped(const int astart, const int asize,
 
 
 #define INT_ROUNDING(result, xrm, gb) \
-  if (gb > 0) { \
-    switch(xrm) {\
+  do { \
+    const uint64_t lsb = 1UL << (gb); \
+    const uint64_t lsb_half = lsb >> 1; \
+    switch (xrm) {\
       case VRM::RNU:\
-        result += ((uint64_t)1 << ((gb) - 1));\
+        result += lsb_half; \
         break;\
       case VRM::RNE:\
-        if ((result & ((uint64_t)0x3 << ((gb) - 1))) == 0x1){\
-            result -= ((uint64_t)1 << ((gb) - 1));\
-            }else if ((result & ((uint64_t)0x3 << ((gb) - 1))) == 0x3){\
-            result += ((uint64_t)1 << ((gb) - 1));\
-        }\
+        if ((result & lsb_half) && ((result & (lsb_half - 1)) || (result & lsb))) \
+          result += lsb; \
         break;\
       case VRM::RDN:\
-        result = (result >> ((gb) - 1)) << ((gb) - 1);\
         break;\
       case VRM::ROD:\
-        result |= ((uint64_t)1ul << (gb)); \
+        if (result & (lsb - 1)) \
+          result |= lsb; \
         break;\
       case VRM::INVALID_RM:\
         assert(true);\
     } \
-  } else if (gb == 0 && xrm == VRM::ROD) { \
-    result |= 1ul; \
-  }
-
+  } while (0)
 
 //
 // vector: integer and masking operand access helper