Incorrect statement in STM32U3, STM32H5, STM32U5 docs

[gbm]

While looking at STM32U3C5 and STM32H503 datasheets I found incorrect information on ICACHE. Quote:

Hit ratio improved by two-ways set-associative architecture and pLRU-t
replacement policy (pseudo-least-recently-used, based on binary tree), algorithm
with best complexity/performance balance

 

The problem is: pseudoLRU may only be applied to 3- (or more) way set-associative caches. With 2-way set-associative, the single flip-flop per set provides true LRU and there is no such thing like “pseudoLRU” for 2 ways. You can’t do binary tree with a single flip-flop.  PseudoLRU is normally used with 4 or more ways, since for 4 ways pseudoLRU requires 3 flip-flops organized as 2-level tree, with simple state transition logic while true LRU - 4 flip-flops and 12 states with much more complex state transition logic.

I suspect the same error might be present in all the STM32 documents describing MCU caches which are 2-way set-associative.

[mƎALLEm]

Hello,

Need to check internally.

I’ll get back to you as soon as I have a feedback.

Internal ticket for follow-up: CDM0063991

[mƎALLEm]

Hello and here is the feedback from the team:

The phrase was meant to express that both elements independently :

• two-ways set-associative architecture
• pLRU replacement policy

Contribute to improve hit ratio

[gbm]

But there is NO psedoLRU in this cache design. PseudoLRU is simply not possible with associativity below 4, so the statement is pure mistake.

[mƎALLEm]

Hello,

Here is the feedback from the team:

“Pseudo-LRU behaves like true LRU when the associativity is below 4.
The statement in the Datasheet : "Hit ratio improved by two-ways set-associative architecture and pLRU-t replacement policy (pseudo-least-recently-used, based on binary tree), algorithm with best complexity/performance balance"
Therefore the statement is generic and it is correct.

Hope that answered your question.

[gbm]

Well, it wasn’t a question, it was an observation.

The bottom line is: it is simply LRU - real, true LRU, not pseudoLRU. So calling it “pseudoLRU” is a mistake. PseudoLRU is used when true LRU cannot be efficiently implemented - typically with 4 or more cache ways. And there is no binary tree of flip-flops with 2-way cache, so that’s yet another mistake.

Of course you may call the light switch in your room “a single-level binary tree of switches, improving the light control using pseudo on/off control” but I believe that “the on/off light switch” name suits and describes it better.

;)

[David Littell]

Give it up, hahaha!  This is the same mindset that decides decades-old cache terminology “flush” is now henceforth to be known as “clean”.  Can’t help ‘em.  🙄