What is the optimal way to sign-extend the lowest n bits of an integer? For example, let’s take a 16-bit integer (the decimal value 34) 0000000000100010 The sign-extend function sext(x,n), with n = 6 should return: 1111111111100010 ^ FEDCBA9876543210 ! bit index In other words, the …

Assuming you understand that the exact bit representation of integer values may change between systems, and that you actually did want the “if the bit at position n-1 is set”, I did it this way. program main use iso_fortran_env, only: int16 implicit none integer(int16), parameter :: to_test…

[image] everythingfunctional: mask = shifta(ibset(0_int16, bit_size(x)-1), bit_size(mask)-n) This creates a mask with the leftmost bits set and the riightmost bits off. There is already a fortran intrinsic maskl() that does exactly this operation. One would need to time the options, or at lea…

I knew there was probably an intrinsic for that. Just didn’t spend enough time looking for it. Edit: to add the updated example. program main use iso_fortran_env, only: int16 implicit none integer(int16), parameter :: to_test = int(b"0000000001100010", int16) print '(b16)', to_test pr…

[image] everythingfunctional: I knew there was probably an intrinsic for that. Just didn’t spend enough time looking for it. Modern fortran now has a pretty complete set of these standard operators for masks, pop counts, parity tests, and so on. This is in contrast to f77 and earlier which ha…

Thanks for spotting the simplification @RonShepard . The shifta(shiftl(...)...) gives reasonably short x86-64 instructions: ; gfortran -O2 -march=x86-64 __extend_MOD_sext: mov ecx, 16 xor eax, eax sub ecx, esi sal edi, cl sar di, cl …

[image] ivanpribec: It translates to directly to the xor instruction as one would expect: Don’t you want ior and not ieor? If any of the left bits are 1, you’ll end up with 0 with ieor.

Oh, you are right. I didn’t consider that using the OR (XOR) like @kargl suggested still needs the branch to check the “sign” bit (position 5 in the function below): ! Sign-extend the lowest n bits of x pure function sext6(x) result(sext) integer(int16), value :: x integer(int16) :: sext if (btest(…

I’ve put the two approaches against each other at the link: Compiler Explorer With both gfortran and ifx the shifting based approach tends to produce between 1 to 4 less instructions. That is both for the n variable case, or the fixed n = 6 case.

[image] ivanpribec: if (btest(x,5)) then sext = ior(x,maskl(10,int16)) else sext = x end if In the bit test fails, do you need to mask out any high-order bits? If so, then something like sext = ibits(x,0,6) or sext = iand( x, maskr(6_int16) ) is indicated. Of these two choices, th…

I am a bit confused. My definition of sign extension is basically what an assign statement in Fortran does when the LHS is a larger kind than the RHS; or that you could do with the INT() function. That is, when storing a signed integer value that is stored in a smaller kind to one with more bits…

How to perform sign-extension?

Help

RonShepard July 14, 2023, 4:42am 3

That is the way I would have done it.

My only suggestion is that the ibits() function in your expression is unnecessary.

sext = shifta(shiftl(x,bit_size(x)-n),bit_size(x)-n)

would work just as well.

Topic		Replies	Views
Moving Bits Question Help	59	2558	November 30, 2022
BITS type in F202Y Language enhancement	27	1660	August 7, 2023
Branchless conditional arithmetic oddities Language enhancement	24	1490	April 7, 2023
Conversion from real to integer Help	30	3746	August 24, 2023
Proper Way to Represent C `unsigned char`? Help	32	2287	October 24, 2024

How to perform sign-extension?

Related topics