fpu: Restrict parts_round_to_int_normal to target precision

Currently parts_round_to_int_normal() assumes that its input has just
been unpacked and therefore doesn't expect non-zero fraction bits past
target precision.

The upcoming DIVIDE TO INTEGER use cases needs it to support
calculations on intermediate values that utilize all fraction bits,
while at the same time restricting the result's precision to frac_size.

Delete the "All integral" check, because even though really large
values are always integer, their low fraction bits still need to be
truncated. For the same reason, make sure rnd_mask covers at least
fraction bits past target precision.

Signed-off-by: Ilya Leoshkevich <iii@linux.ibm.com>
Reviewed-by: Thomas Huth <thuth@redhat.com>
Message-ID: <20260210214044.1174699-4-iii@linux.ibm.com>
Signed-off-by: Thomas Huth <thuth@redhat.com>

diff --git a/fpu/softfloat-parts.c.inc b/fpu/softfloat-parts.c.inc
index 5e0438fc0b7f8fbfb9d0a27747188e84131f2de9..79b56014ab9887bfba6ae2c699311e9087b5f2dc 100644 (file)
--- a/fpu/softfloat-parts.c.inc
+++ b/fpu/softfloat-parts.c.inc
@@ -1118,11 +1118,6 @@ static bool partsN(round_to_int_normal)(FloatPartsN *a, FloatRoundMode rmode,
         return true;
     }
 
-    if (a->exp >= frac_size) {
-        /* All integral */
-        return false;
-    }
-
     if (N > 64 && a->exp < N - 64) {
         /*
          * Rounding is not in the low word -- shift lsb to bit 2,
@@ -1132,8 +1127,13 @@ static bool partsN(round_to_int_normal)(FloatPartsN *a, FloatRoundMode rmode,
         frac_shrjam(a, shift_adj);
         frac_lsb = 1 << 2;
     } else {
+        /*
+         * Rounding is in the low word -- compute the lsb offset for rounding
+         * and for clamping to the target precision, then map it to an offset
+         * within frac_lo.
+         */
         shift_adj = 0;
-        frac_lsb = DECOMPOSED_IMPLICIT_BIT >> (a->exp & 63);
+        frac_lsb = DECOMPOSED_IMPLICIT_BIT >> (MIN(a->exp, frac_size) & 63);
     }
 
     frac_lsbm1 = frac_lsb >> 1;