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/vnclip_vv.h b/riscv/insns/vnclip_vv.h
index 7bcb4cb..4ec0f40 100644
--- a/riscv/insns/vnclip_vv.h
+++ b/riscv/insns/vnclip_vv.h
@@ -4,20 +4,15 @@ int64_t int_max = (1 << (P.VU.vsew - 1)) - 1;
 int64_t int_min = -(1 << (P.VU.vsew - 1));
 VI_VVXI_LOOP_NARROW
 ({
+  int128_t result = vs2;
+  unsigned shift = vs1 & ((sew * 2) - 1);
 
-  int64_t result = vs2;
-// rounding
-  INT_ROUNDING(result, xrm, sew);
+  // rounding
+  INT_ROUNDING(result, xrm, shift);
 
-// unsigned shifting to rs1
-  uint64_t unsigned_shift_amount = (uint64_t)(vs1 & ((sew * 2) - 1));
-  if (unsigned_shift_amount >= (2 * sew)) {
-    unsigned_shift_amount = 2 * sew - 1;
-  }
-
-  result = (vsext(result, sew * 2)) >> unsigned_shift_amount;
+  result = result >> shift;
 
-// saturation
+  // saturation
   if (result < int_min) {
     result = int_min;
     P.VU.vxsat = 1;