Re: [PATCH] RISC-V: Add the Zihpm and Zicntr extensions

[Jeff Law <jeffreyalaw@gmail.com>]

On 11/22/22 08:29, Palmer Dabbelt wrote:
> On Tue, 22 Nov 2022 07:20:15 PST (-0800), jeffreyalaw@gmail.com wrote:
>>
>> On 11/20/22 18:36, Kito Cheng wrote:
>>>> So the idea here is just to define the extension so that it gets 
>>>> defined
>>>> in the ISA strings and passed through to the assembler, right?
>>> That will also define arch test marco:
>>>
>>> https://github.com/riscv-non-isa/riscv-c-api-doc/blob/master/riscv-c-api.md#architecture-extension-test-macro 
>>>
>>
>> Sorry I should have been clearer and included the test macro(s) as well.
>>
>> So a better summary would be that while it doesn't change the codegen
>> behavior in the compiler, it does provide the mechanisms to pass along
>> isa strings to other tools such as the assembler and signal via the test
>> macros that this extension is available.
>
> IMO the important bit here is that we're not adding any compatibility 
> flags, like we did when fence.i was removed from the ISA.  That's fine 
> as long as we never remove these instructions from the base ISA in the 
> software, but that's what's suggested by Andrew in the post.

Right.  IIUC these instructions were never supposed to be in the base 
ISA, but, in effect, snuck through.  We're retro-actively adding them as 
an extension, at least in terms of ISA strings & test macros.  We're 
currently (forever?) going to allow them in the assembler without 
strictly requiring the extension be on.


> It's a super weird one, but there's a bunch of cases in RISC-V where 
> we're told to just ignore words in the ISA manual.  Definitely a trap 
> for users (and we already had some Linux folks get bit by the counter 
> changes here), but that's just how RISC-V works.

Mistakes happen.  The key is to adjust for them as best as we can.    
I'd lean towards a stricter enforcement, bringing these 
instructions/extension in line with how we handle the others. It'd 
potentially mean source incompatibilities that would need to be fixed, 
but they shouldn't be difficult and we're still early enough in the game 
that we *could* take that route.  Andrew's position is more 
accommodating of existing code and while I may not 100% agree with his 
position, I understand it.


So while I'd lean towards a stricter checking, I can live with this 
approach.  I wouldn't mind hearing from Kito, Philipp and others though.


Jeff