rvv: fix the rounding bit position for vnclip instructions.

1. The rounding increment should be derived from the shift amount, not SEW.
2. Use 128bit to store temporary result to handle shift = 63 case in rv64

Signed-off-by: Albert Ou <aou@eecs.berkeley.edu>
Signed-off-by: Chih-Min Chao <chihmin.chao@sifive.com>

1 files changed, 6 insertions, 11 deletions

diff --git a/riscv/insns/vnclipu_vx.h b/riscv/insns/vnclipu_vx.h
index 0507a2b..5b064ce 100644
--- a/riscv/insns/vnclipu_vx.h
+++ b/riscv/insns/vnclipu_vx.h
@@ -3,20 +3,15 @@ VRM xrm = P.VU.get_vround_mode();
 uint64_t int_max = ~(-1ll << P.VU.vsew);
 VI_VVXI_LOOP_NARROW
 ({
-  uint64_t result = vs2;
+  uint128_t result = vs2_u;
+  unsigned shift = rs1 & ((sew * 2) - 1);
 
-// rounding
-  INT_ROUNDING(result, xrm, sew);
+  // rounding
+  INT_ROUNDING(result, xrm, shift);
 
-// unsigned shifting to rs1
-  uint64_t unsigned_shift_amount = (uint64_t)(rs1 & ((sew * 2) - 1));
-  if (unsigned_shift_amount >= (2 * sew)) {
-    result = 0;
-  } else {
-    result = vzext(result, sew * 2) >> unsigned_shift_amount;
-  }
+  result = result >> shift;
 
-// saturation
+  // saturation
   if (result & (uint64_t)(-1ll << sew)) {
     result = int_max;
     P.VU.vxsat = 1;