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, 5 deletions

diff --git a/riscv/insns/vnclip_vi.h b/riscv/insns/vnclip_vi.h
index ca27593..287035e 100644
--- a/riscv/insns/vnclip_vi.h
+++ b/riscv/insns/vnclip_vi.h
@@ -4,14 +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
 ({
-
   int64_t result = vs2;
-// rounding
-  INT_ROUNDING(result, xrm, sew);
+  unsigned shift = zimm5 & ((sew * 2) - 1);
+
+  // rounding
+  INT_ROUNDING(result, xrm, shift);
 
-  result = vsext(result, sew * 2) >> (zimm5 & ((sew * 2) < 32? (sew * 2) - 1: 31));
+  result = result >> shift;
 
-// saturation
+  // saturation
   if (result < int_min) {
     result = int_min;
     P.VU.vxsat = 1;