Underflow/Overflow in mul_2pow

[damianmoz]

I found reading this description clunky.

Overflow will produce infinity. The result will be zero rather than a
subnormal number in case of underflow, regardless of control bits in
the mask or numeric control register.

Can I suggest changing the second sentence to match the structure of the first:

Overflow produces infinity. Underflow produces zero rather than a
subnormal number in case of underflow, regardless of control bits in
the mask or numeric control register.

Does this means that if I have code like

let x = 0x1.0p-1022 * 0.25

that I will get a value of 'x' which is zero? Probably not what I would prefer.
Or, if the use of that instruction intended to be dependent on compiler options.
Or have I misunderstood things.

Is there an instruction which behaves as IEEE 754's 'scalbn'

[agner]

The main purpose of the mul_2pow instruction is to improve speed. x * 0.25 will take several clock cycles. With optimization on, the compiler or assembler can change x*0.25 to mul_2pow(x, -2) which takes only 1 clock cycle. It would be difficult to do this in a single clock cycle if you want underflow to produce a subnormal number.

Apart from not supporting subnormal numbers, I think this instruction behaves like scalbn. Can you spot any difference?

[damianmoz]

Understand the reasoning for speed. If I am ever lucky enough to have hardware based on this instruction set, I will just have to be careful to protect myself if I want to avoid flushing to zero. I was curious if you could use a template with an option to enable/disable such flushing to zero.

And yes, the instruction behaves the same as scalbn except for subnormal results.