aboutsummaryrefslogtreecommitdiff
diff options
context:
space:
mode:
-rw-r--r--fpu/softfloat.c6
1 files changed, 6 insertions, 0 deletions
diff --git a/fpu/softfloat.c b/fpu/softfloat.c
index d90d79d777..70e0c40a1c 100644
--- a/fpu/softfloat.c
+++ b/fpu/softfloat.c
@@ -1878,6 +1878,12 @@ static FloatParts scalbn_decomposed(FloatParts a, int n, float_status *s)
return return_nan(a, s);
}
if (a.cls == float_class_normal) {
+ /* The largest float type (even though not supported by FloatParts)
+ * is float128, which has a 15 bit exponent. Bounding N to 16 bits
+ * still allows rounding to infinity, without allowing overflow
+ * within the int32_t that backs FloatParts.exp.
+ */
+ n = MIN(MAX(n, -0x10000), 0x10000);
a.exp += n;
}
return a;