Cmpsb multithread?

Index > High Level Languages > Cmpsb multithread?

Author

Thread

l4m2

Joined: 15 Jan 2015
Posts: 670

l4m2 19 Oct 2020, 17:18

Code:

#include <stdio.h>
#include <thread>
volatile int x;
void p1() {
    int ret;
    while (1) {
        asm("cmpsb;pushf;pop %%rax":"=A"(ret):"D"(&x), "S"(&x));
        if(ret&(1<<6)) { // Can't compile setz
            putchar(' ');
        } else {
            putchar('*');
        }
    }
}
int main() {
    std::thread p(p1);
    while (1) {
        x = 0;
        x = -1;
    }
}

Running this puts some *.
1. On same core different thread, there're lots of *s; on different core, less *s; on same thread, no *. So core's cache can be updated during an instruction?
2. document only says

Code:

temp ← SRC1 - SRC2;
SetStatusFlags(temp);

Is it a promise that SRC1 is read first?

19 Oct 2020, 17:18

bitRAKE

Joined: 21 Jul 2003
Posts: 4249
Location: vpcmpistri

bitRAKE 19 Oct 2020, 19:02

I built it with clang and didn't get a single "*".
Kind of curious what you're running on?
Running on a Ryzen 2700, atm.

_________________
¯\(°_o)/¯ “languages are not safe - uses can be” Bjarne Stroustrup

19 Oct 2020, 19:02

l4m2

Joined: 15 Jan 2015
Posts: 670

l4m2 19 Oct 2020, 19:22

I'm on i7-4790

19 Oct 2020, 19:22

bitRAKE

Joined: 21 Jul 2003
Posts: 4249
Location: vpcmpistri

bitRAKE 19 Oct 2020, 22:59

I didn't think it was possible to split the update but tried to catch a problem with multiple threads. Still nothing.

If this program returns zero on your machine then your results are coming from somewhere else.

Edit: whoops I didn't change anything, lol.

Edit: this one catches it happening - definitely surprised. 17586658 splits out of 15*2^32.

Description:		Download
Filename:	cmpsb.zip
Filesize:	2.73 KB
Downloaded:	732 Time(s)

_________________
¯\(°_o)/¯ “languages are not safe - uses can be” Bjarne Stroustrup

19 Oct 2020, 22:59

bitRAKE

Joined: 21 Jul 2003
Posts: 4249
Location: vpcmpistri

bitRAKE 20 Oct 2020, 01:02

I'm sorry about the subsystem version, but I don't have an old linker on this machine. It doesn't use anything fancy - should probably run fine on XP.

20 Oct 2020, 01:02

revolution
When all else fails, read the source

Joined: 24 Aug 2004
Posts: 20667
Location: In your JS exploiting you and your system

revolution 20 Oct 2020, 03:42

What does the addition of lock do?

20 Oct 2020, 03:42

bitRAKE

Joined: 21 Jul 2003
Posts: 4249
Location: vpcmpistri

bitRAKE 20 Oct 2020, 05:49

I've added the option to toggle a single byte and added some comments about performance: https://github.com/bitRAKE/fasmg_playground/blob/master/win64.coff/cmpsb_split.g

Adding a lock to the data would obviously prevent the data from changing during another thread access. The downside is that I'd have 15 threads lining up for access to that data and the execution time goes up substantially. The main goal here was to show that the two reads of CMPSB can be split. It's not exactly a practical use-case.

If you're inquiring about the LOCK ADD existing in the code - that is to ensure the counts are correct and not to synchronize anything else.

My processor can hide the collisions if there is other bus traffic - that is why the threads need to be high-priority to force collisions. I'm impressed with how well it does.

_________________
¯\(°_o)/¯ “languages are not safe - uses can be” Bjarne Stroustrup

20 Oct 2020, 05:49

revolution
When all else fails, read the source

Joined: 24 Aug 2004
Posts: 20667
Location: In your JS exploiting you and your system

revolution 20 Oct 2020, 06:05

I think that what is really being tested here is the CPU data access patterns.

I see no reason why the CPU should ensure the two reads by cmpsb are protected from another thread's access. It is always the programmers responsibility to ensure threads don't stomp on other's data.

20 Oct 2020, 06:05

bitRAKE

Joined: 21 Jul 2003
Posts: 4249
Location: vpcmpistri

bitRAKE 20 Oct 2020, 06:35

I agree. Could even compare it to two basic instructions and see if it has the same profile. Or we could see if all processors perform the two reads in the same order.

20 Oct 2020, 06:35

revolution
When all else fails, read the source

Joined: 24 Aug 2004
Posts: 20667
Location: In your JS exploiting you and your system

revolution 20 Oct 2020, 06:45

bitRAKE wrote:

Adding a lock to the data would obviously prevent the data from changing during another thread access.

I meant to try lock cmpsb.

20 Oct 2020, 06:45

bitRAKE

Joined: 21 Jul 2003
Posts: 4249
Location: vpcmpistri

bitRAKE 20 Oct 2020, 07:17

Illegal instruction. Razz

_________________
¯\(°_o)/¯ “languages are not safe - uses can be” Bjarne Stroustrup

20 Oct 2020, 07:17

l4m2

Joined: 15 Jan 2015
Posts: 670

l4m2 20 Oct 2020, 16:42

Changed changer loop to

Code:

lock incl (x)

CY is always 0 unless incl carry.

PS. Why High Level Lang? C++ part only serve putchar and thread

20 Oct 2020, 16:42

revolution
When all else fails, read the source

Joined: 24 Aug 2004
Posts: 20667
Location: In your JS exploiting you and your system

revolution 20 Oct 2020, 23:13

l4m2 wrote:

PS. Why High Level Lang? C++ part only serve putchar and thread

Because we need an HLL to compile your code. It doesn't involve fasm, so it doesn't belong in Main where you posted it originally.

20 Oct 2020, 23:13

< Last Thread | Next Thread >

Forum Rules:

You cannot post new topics in this forum
You cannot reply to topics in this forum
You cannot edit your posts in this forum
You cannot delete your posts in this forum
You cannot vote in polls in this forum
You cannot attach files in this forum
You can download files in this forum