Pascal - inc -> assembler - add?
lazarus / 62-General

----------------------------------------------------------------------
Seenkao · 2021-08-06 18:16

Code: Pascal  [Select][+][-]window.onload = function(){var x1 =
document.getElementById("main_content_section"); if (x1) { var x =
document.getElementsByClassName("geshi");for (var i = 0; i begin  i :=
0;  inc(i);        // to assembler: add $0x1, %r14dend;
Почему не используется инструкция ассемблера INC? Если это невозможно
для данного регистра, почему не используется один из основных
регистров?

google translate:
Why isn't the INC assembler instruction being used? If this is not
possible for a given register, why is one of the main registers not
being used?

----------------------------------------------------------------------
MathMan · 2021-08-07 08:36

Hello Seenkao,

I need to speculate a bit because I am not involved in FPC compiler
code generation development, but

* not all processor architectures have an 'inc' instruction
* on modern x86-64 CPUs 'inc' is not faster/better than 'add' <=
except some very rare and tricky corner cases
* in FPC you can also do e.g. 'Inc( i, 5 )', which can only be
translated to an 'add' instruction anyway

In combination the above will lead to a much simpler code generation
if asm 'add' is used for Pascal 'inc()' throughout, i guess.

Cheers,
MathMan

----------------------------------------------------------------------
Kays · 2021-08-07 11:43

Quote from: Seenkao on August 06, 2021, 08:16:07 pm[…] Why isn't the
INC assembler instruction being used? […]Potential partial flags
stall.
Quote from: Seenkao on August 06, 2021, 08:16:07 pm[…] If this is not
possible for a given register, why is one of the main registers not
being used?That’s how {$optimization regVar} is implemented. It would
increase the compiler’s complexity if you wanted to use, for instance,
rax for that purpose, yet the gain in performance is minimal/non-
existent. Certainly you will need the REX prefix for any instruction
using one of the r8–r15 registers, but the underlying operation (here
add) performs just as fast, you know.

----------------------------------------------------------------------
Leledumbo · 2021-08-07 12:48

You want inc and all of the consequences written by Kays? Simple,
optimize for size with -Os.

----------------------------------------------------------------------
PascalDragon · 2021-08-07 15:31

Quote from: Seenkao on August 06, 2021, 08:16:07 pmWhy isn't the INC
assembler instruction being used? If this is not possible for a given
register, why is one of the main registers not being used?

To quote from the compiler's sources (compiler/x86/cgx86.pas, line
197):

Code: Pascal  [Select][+][-]window.onload = function(){var x1 =
document.getElementById("main_content_section"); if (x1) { var x =
document.getElementsByClassName("geshi");for (var i = 0; i     {
modern CPUs prefer add/sub over inc/dec because add/sub break
instructions dependencies on flags      because they modify all flags
}

----------------------------------------------------------------------
440bx · 2021-08-07 15:55

The number of subtleties involved in generating and optimizing code
are frightening.

It's much too easy to make incorrect, or at least questionable
assumptions, about the behavior of a series of instructions.

----------------------------------------------------------------------
Seenkao · 2021-08-07 17:38

Благодарю за информацию!
Давно не занимался ассемблером... надо обновить знания.

Google translate:
Thanks for the information!
I haven't studied assembler for a long time ... I need to update my
knowledge.

----------------------------------------------------------------------
MathMan · 2021-08-09 09:35

Quote from: 440bx on August 07, 2021, 05:55:45 pmThe number of
subtleties involved in generating and optimizing code are frightening.

It's much too easy to make incorrect, or at least questionable
assumptions, about the behavior of a series of instructions.

And it only starts with the code generated by the compiler. Each
hardware implementation of an architecture (e.g. x86-64 instruction
set) has subtle quirks that are not visible from the outside.

Since some time there are "superoptimizers" <= programs that take as
input a sequence of assembler instructions (generated by hand or out
of a compiler), then shuffle the instructions within given limits
(maintaining execution results) and finally measure execution speed of
the different variants. On modern CPU (e.g. Intel Skylake) one can
regularly see speed improvements of ~5%.

But even these "superoptimizers" will not achieve optimum, because
they are bound by the instructions given in the input. I can show you
examples, where I achieved a speed improvement of ~20% by inserting a
"useless" AVX2 instruction into a core loop which otherwise only uses
general purpose registers. Nobody could explain this to me so far -
and yes, I did address this to Intel directly too.

Cheers,
MathMan

----------------------------------------------------------------------
JuhaManninen · 2021-08-09 14:28

Interesting.
I haven't done assembly programming since 1980's but I follow CPU
architecture innovations.
I learned the details of RISC-V design. Wow, just amazing! A true RISC
architecture with optional extensions. Support for modern optimization
techniques are thought of and built in the design.
This has nothing to do with Pascal, thus possible further discussion
should happen in the "Other" section.
Still, amazing.