Linux Kernel Set To Finally Retire AMD 3DNow!

[jabl]

For a while, 3DNow! was referenced as "Integer SSE", eventhough it's false - both SSE and 3DNow! could process 64-bit FP values. SSE could do pretty much the same, but also added the ability to process 128-bit values.

Neither SSE nor 3dnow are able to work with 64-bit FP values. That only came with SSE2. 3dnow packed two single precision (32-bit) floating point values into a single register (the 3dnow register file was aliased with the x87 registers, so you couldn't use both at the same time), and SSE introduced the 128-bit xmm registers containing 4 32-bit FP values.

[rmfx]

So we add features with hundreds of KLOC, but we remove a small chunk of well tested code because...progress?

At least they didn't remove support for these processors entirely.

Yep, there are millions of other bs lines of code to remove. Like the graphic drivers. A light and universal cpu vulkan support a la lavapipe should be enough to drive graphics until mesa or blobs are installed.

[tildearrow]

I guess time to upgrade my server machine already...

[mitch074]

That is false. Read the PDF documentation (the link is below).

The main point of SSE is that it added separate 128-bit registers. x87, MMX and 3DNow share the same 64-bit registers.
https://www.amd.com/system/files/TechDocs/21928.pdf

FP64 : true, sorry. It could process 2 FP32 values in a single instruction,which doesn't make it a FP64 unit.
Registers : SSE added 8 more registers while 3DNow! shared the 8 existing MMX registers, true - and it's a GOOD THING they did, due to how starved for registers the x86-32 arch is... And considering Intel created SSE to prevent 3DNow! from becoming prevalent (as it had been adopted by all its x86-32 competitors), they had to make it better. AMD had already done good with 3DNow! by allowing simultaneous processing of 2 FP instructions in a single cycle, SSE had to be better everywhere.

[baryluk]

BTW. This just drops use of 3DNow! from kernel to accelerate memory copies and comparison. SSE, SSE2, AVX are available for about 25 years now.

You can still run user space applications using 3DNow! (including native and via wine) just fine on CPUs that support it.

[ResponseWriter]

Do you need 3DNow? Does anything use it, really?

You could argue that about a lot of things. I don't, for example, use or need AVX512. Like I said, the hardware supports SSE4a so for me the loss of 3DNow! support will hardly be noticed.

[Nuc!eoN]

With the global chip shortage, and no end in sight it is beyond me why support/features for older hw generations is getting removed from the linux kernel and mesa recently.

[pal666]

Whoever wants to limit the waste he produces will use its tools till they can't be repaired. Buying something new and throwing away what still works causes more energy consumption and pollution.

unless that new is much less power hungry(which it will be if new enough)

[tildearrow]

That might be true if money is your only concern. Whoever wants to limit the waste he produces will use its tools till they can't be repaired. Buying something new and throwing away what still works causes more energy consumption and pollution.

My home NAS has 30 - 35W idle (old Phenom2 quad core) but it can deliver > 100 MB/s filetransfers with full disk encription on raid 5. I have a Raspi3 that i tested to replace it, but its NAS capabilities are weak ~ 10MB/s without raid and encription. Thats to slow if you really need a NAS, not to speak of multiple people using it at the same time.

But here my Phenom is as slow as a Raspberry Pi 4B... Yep, 50MB/s over SFTP...

[jabl]

So we add features with hundreds of KLOC, but we remove a small chunk of well tested code because...progress?

Evidently the code wasn't that well tested since it was broken. From the commit message:

This code puts an exception table entry on the PREFETCH instruction to overwrite it with a JMP.d8 when it triggers an exception. Except of course, our code is no longer writable, also SMP.

Instead of fixing this broken mess, simply take it out.