Motorola DSP96002 User Manual page 773

D.1.5.1.2.2 Results Rounded To SP That Are Normalized
If the Data ALU operation result was rounded to SP and the rounded result may be represented as a nor-
malized single precision floating-point number, the result will be stored in normalized DP format that may
be read out by single and double precision moves without errors or truncation.
D.1.5.1.2.3 Results Rounded To SP That Are Denormalized
If the Data ALU operation result was rounded to SP and the rounded result must be represented as a de-
normalized single precision floating-point number, the result will be stored in unnormalized DP format with
the U tag set and the I bit cleared, and it may be read out by single precision moves without errors or trun-
cation. If the register is read by a double precision move, completely incorrect data will be obtained; see
the discussion in Section C.3.2.
In this case, before the result is delivered, an additional Data ALU execution cycle is required in which the
SEP mantissa is shifted right the required number of places for correct rounding to SP.
The presence of unnormalized numbers in DP format will add one dummy cycle followed by an additional
cycle for each unnormalized DP operand to any Data ALU operation that uses them as input. During the
additional cycle the unnormalized operand (U-TAG=1) is normalized, however the register itself will not be
modified.
D.1.5.1.2.4 Results Rounded To SEP
Data ALU results are rounded to SEP when the instruction is specified with the .X suffix (FMPY.X, FADD.X,
etc.).
D.1.5.1.2.5 Results Rounded To SEP That Are Normalized
If the Data ALU operation result was rounded to SEP and the rounded result may be represented as a nor-
malized single extended precision floating-point number, the result will be stored in normalized DP format
that may be read out by double precision moves without errors or truncation.
If the result stored in the register is read with a single precision move, two situations may occur:
The SEP exponent is in the range of the normalized SP exponent: the data read will be round-
1.
ed to SP by truncating the SEP mantissa; this is equivalent to IEEE round towards zero.
The SEP exponent is not in the range of the normalized SP exponent: the data read will not
2.
have the right exponent. The correct value should have been infinity, zero or a denormalized
SP, but the move instruction does not provide it.
D.1.5.1.2.6 Results Rounded To SEP That Are Denormalized
If the Data ALU operation result was rounded to SEP and the rounded result must be represented as a
denormalized single extended precision floating-point number, the result will be stored in normalized DP
format with the V tag set and I bit cleared, and it may be read out by double precision
moves without errors or truncation. If the register is read by a single precision move, completely incorrect
data will be obtained; see the discussion in Section C.3.3 (double precision and single extended precision
numbers have the same exponent bias).
MOTOROLA DSP96002 USER'S MANUAL D-21