Clamping (Limit) Functions

Although not specified in any IEEE 754 standard, the ISA supports a clamping (or limit) instruction (FFLIM) intended for use where an input operand needs to be constrained between an upper and lower limit. It serves a similar purpose to the integer equivalent FLIM instruction and is essentially a concurrent execution of FMIN and FMAX operations with a common operand. Refer to the truth table shown in Table 4-12 for a definition of how NaN operands are handled.

Any finite operand value will compare as less than +infinity or greater than -infinity. An operand value of -0 compares to less than +0.

For FFLIM operations, when both upper and lower limits are either both qNaN or both sNaN values, a NaN significant comparison is not required, and the Fb NaN will be the default source for the result. This differs from how coincident NaN values are treated in general.

Furthermore, a NaN input value (Fd) will cause the limit values to be ignored and will become the source for the result. That is, checks between input NaNs and limit values (NaNs or otherwise) are also not required.

Table 4-12. FFLIM Operation⁽⁴⁾
Fb⁽³⁾ (Lower Limit)	Fs⁽³⁾ (Upper Limit)	Fd (Input Value)	Invalid Exception Mask	Fd (Result)	FSR.INVAL	Exception Taken?
FPNL⁽¹⁾	FPNU	FPN	Don’t care	FPNL or FPNU or FPN⁽²⁾ or Distinguished qNaN⁽³⁾	0	No
FPNL	qNaN_U	FPN	Don’t care	qNaN_U	0	No
FPNL	sNaN_U	FPN	1	Quieted sNaN_U	1	No
FPNL	sNaN_U	FPN	0	Quieted sNaN_U	1	Yes
qNaN_L	FPNU	FPN	Don’t care	qNaN_L	0	No
sNaN_L	FPNU	FPN	1	Quieted sNaN_L	1	No
sNaN_L	FPNU	FPN	0	Quieted sNaN_L	1	Yes
sNaN_L	sNaN_U	FPN	1	Quieted sNaN_L⁽⁵⁾	1	No
sNaN_L	sNaN_U	FPN	0	Quieted sNaN_L⁽⁵⁾	1	Yes
sNaN_L	qNaN_U	FPN	1	qNaN_U⁽⁶⁾	1	No
sNaN_L	qNaN_U	FPN	0	qNaN_U⁽⁶⁾	1	Yes
qNaN_L	sNaN_U	FPN	1	qNaN_L⁽⁶⁾	1	No
qNaN_L	sNaN_U	FPN	0	qNaN_L⁽⁶⁾	1	Yes
qNaN_L	qNaN_U	FPN	Don’t care	qNaN_L⁽⁵⁾	0	No
Don’t care	Don’t care	sNaN	1	Quieted sNaN⁽⁷⁾	1	No
Don’t care	Don’t care	sNaN	0	Quieted sNaN⁽⁷⁾	1	Yes
Don’t care	Don’t care	qNaN	Don’t care	qNaN	0	No
Note: FPNL and FPNU are floating-point numbers that are not a NaN. Result determined by FFLIM operation. If Fs is less than Fb (and neither Fs nor Fb are NaN values), the result will be the distinguished qNaN, and the invalid exception will be signaled. FFLIM operation based on IEEE 754-2019 minimum(x,y) and maximum(x,y) operation definitions. Unlike FMIN/FMAX operations, no magnitude comparison of limit NaN values is required. The default result will always be sourced from Fb. qNaN values have priority over sNaN values (see Table 4-13). Unlike FMIN/FMAX operations, no comparison of limit and input (Fd) values are required. The default result will always be sourced from Fd.