Re: Top 10 most common hard skills listed on resumes...

On 12/09/2024 23:28, Waldek Hebisch wrote:

David Brown <david.brown@hesbynett.no> wrote:

On 12/09/2024 01:59, Waldek Hebisch wrote:

David Brown <david.brown@hesbynett.no> wrote:

>
On many cpus, using sizes smaller than the full register size means
doing sign extensions or masking operations at various times - thus full
size register operations can often be more efficient.  On such systems
you will find that int_fast16_t is 32-bit or 64-bit, according to the
register width.  On other cpus, some common ALU operations on full-size
operands can be slower than for smaller operands (such as on the 68000).
   There, int_fast16_t will be 16-bit.
>
Compiler authors know what will usually be faster on the target.  There
will always be some exceptions (division is usually faster on smaller
operands, for example).  But if you don't know the target - as is the
case of portable code - the compiler will usually make a better choice
here than you would.

>
BTW, I just played with Clang 18 on 64-bit FreeBSD.  It has 32-bit
int_fast16_t.  gcc in Linux makes it 64-bit.  Who is right?
>

>
Technically, both are right - implementations can use any integer type
of at least 16 bits here, whatever they think is fastest in general.
But it surprises me somewhat, given that clang for x86-64 on Linux uses
64-bit for int_fast16_t.

 Well, both satisfy "hard" requirements.  But the question was which
type is faster.

Yes.  But unless I am using a target processor that I know well, have a good idea of the types of instructions and know about the timings for those instructions with different data types, then I am inclined to believe the compiler implementer here and use the int_fastNN_t types. (For most of my C programming, I /do/ know the target well, and can make more refined type choices if it is relevant.  But I don't know the timing details for the countless x86-64 variants.)

 

But to be clear, the size of the "fast" types depends on the target and
the implementation.  They are not normally used for external ABI's, and
are purely internal to the generated code.  Obviously you must pick a
"fast" size that is at least as big as the range you need.

 I think that Linux (and probably FreeBSD too) considers size of
fast type as part of ABI (regardless of gudelines those types
certainly leaked into "public" interfaces).  Such ABI change is
probably viewed as not worth doing.
 

I am not sure if the fast types are in the ABI.  It certainly seems not, if you say clang on x86-64 BSD has 32-bit int_fast32_t, while it is 64-bit on Linux gcc and clang.  BSD and Linux use the same ABI, AFAIK. (Everybody except MS use the same x86-64 ABI.)

And concering choice on x86_64, AFAIK for operations on numbers of
the same size 32-bit gives fastest operations.  16-bit had two
disadvantages, big one due to partial register stalls, small one
due to larger size (operand size prefix).  64-bit requires bigger
code (more need to use prefixes) and bigger data.

I don't know if that is all correct or not.  Some operations are definitely slower on bigger operands, such as division.  Different x86-64 processors may see different costs for things like prefix sizes. And if you can get SIMD instructions into the picture, then smaller sizes let you do more in the same instruction.

 When mixing
types, 32-bit numbers are automatically zero extended, so there
is no extra cost when mixing unsigend numbers.

When I look at generated code, unsigned types smaller than 64 bits can require masking at times, and they do sometimes require zero extend instructions (typically expressed as a "move 32-bit register to 64-bit register" instruction).

 So what remains
is mixing signed 32-bit integers with 64-bit ones.  Addresses
use 64-bit artitmetic, so that requires sign extention.  OTOH
in arithmetic "fast" types are likely to be mixed with exact
32-bit types and then making "fast" types 32-bit is faster
overall.  So, there are bets/assumptions which usage is more
frequent.  OTOH, choice between 32-bit and 64-bit fast types
is unlikely to make _much_ difference.
 

It could be interesting to compare speeds for different kinds of code on different x86-86 targets.  But the details here are beyond me, and also well outside of the targets that I am most interested in.