/g/ - Thread 108599109

>but chud branch predictor, you're accessing memory at an address that's a function of privileged data without bound checking!
>If something happens, you'll load a specific byte to cache which the current thread can extract via a timing attack, inferring the contents of the privileged memory!

Anonymous
04/14/26(Tue)00:54:54 No.108601360

Anonymous 04/14/26(Tue)00:54:54 No.108601360▶

kek

Anonymous
04/14/26(Tue)01:10:57 No.108601448

Anonymous 04/14/26(Tue)01:10:57 No.108601448▶

File: 1743628183976710.jpg (105 KB)

105 KB JPG

>>108599109
The Linux kernel code is full of these
if (unlikely(ptr))
    return 0;

Anonymous
04/14/26(Tue)01:25:13 No.108601534

Anonymous 04/14/26(Tue)01:25:13 No.108601534▶

>>108599109
That's literally the ryzen platform.
There's a mini neural network chud that decides it

Anonymous
04/14/26(Tue)01:27:57 No.108601550

Anonymous 04/14/26(Tue)01:27:57 No.108601550▶

>>108601534
Good self-modifying code doesn’t need a fucking branch predictor.
Do likewise, gents.

Anonymous
04/14/26(Tue)01:36:33 No.108601604

Anonymous 04/14/26(Tue)01:36:33 No.108601604▶

>>108599109
nothing ever jumps forward

Anonymous
04/14/26(Tue)01:40:49 No.108601619

Anonymous 04/14/26(Tue)01:40:49 No.108601619▶

mitigations=auto nosmt=force memtest=4

Anonymous
04/14/26(Tue)01:41:36 No.108601622

Anonymous 04/14/26(Tue)01:41:36 No.108601622▶

>>108601550
>self modifying code
>*incessantly invalidates your instruction cache*
what do?

Anonymous
04/14/26(Tue)02:06:21 No.108601730

Anonymous 04/14/26(Tue)02:06:21 No.108601730▶

>>108601622
What? just run as it is on the cache, it's roughly the same shit so it's no problem!

Anonymous
04/14/26(Tue)02:19:16 No.108601785

Anonymous 04/14/26(Tue)02:19:16 No.108601785▶

mutts won, humans lost

Anonymous
04/14/26(Tue)02:45:19 No.108601897

Anonymous 04/14/26(Tue)02:45:19 No.108601897▶

>>108601730
If you use self modifying code to skip branches, then I assume you mean removing branches from the machine code in memory, right?
The problem is that this invalidates the instruction cache line for the branch, which will cause a far worse stall than a mispredicted branch.
If you determine at runtime that your code always or never takes a branch, then it's probably marginally better to pay the stall penalty once, rather than use up space in the branch predictor.
If your branch is sometimes taken, then it's basically never worth it to modify the instruction on that condition. The branch predictor will essentially do that for you in a much faster way.

Anonymous
04/14/26(Tue)02:52:08 No.108601924

Anonymous 04/14/26(Tue)02:52:08 No.108601924▶

>>108601897
I wasnt making a that smart post, just a "stupid CPU execute old version of code" deal that fucks with everything.

Anonymous
04/14/26(Tue)03:11:36 No.108602000

Anonymous 04/14/26(Tue)03:11:36 No.108602000▶

>>108601622
You backpatch at least 64 bytes behind if you’re worried about your precious cache line.
> my code fits in a cache line
Good, but unroll the first iteration of it. The first iteration patches the next before the prefetch, and the subsequent iteration(s) patch the first or previous unrolled iteration.

Anonymous
04/14/26(Tue)03:39:18 No.108602115

Anonymous 04/14/26(Tue)03:39:18 No.108602115▶

File: 1775926991428056.jpg (11.1 KB)