Use better types for virtual memory translation code #637
Conversation
|
I wrote this code quite a while ago and kind of got sidetracked by other stuff, but it looks like I finished it so I thought I may as well push a draft and see what people think. It hasn't been tested extensively yet but I can do that. One slightly awkward thing I found is that you can use really nice strict types everywhere for the actual translation because the lifetime of |
Timmmm
force-pushed
the
user/timh/ptw_nice_types
branch
from
55e080e to
8c15f7c
Compare
Timmmm
force-pushed
the
user/timh/ptw_nice_types
branch
from
8c15f7c to
dcb8d5b
Compare
|
Ok I think this is ready. I will run it through our CI system and report the test results. |
Timmmm
force-pushed
the
user/timh/ptw_nice_types
branch
from
dcb8d5b to
079b1a7
Compare
Timmmm
force-pushed
the
user/timh/ptw_nice_types
branch
from
559d67b to
4063c08
Compare
Change the VM code from using `bits(64)` everywhere to using more accurate types. This also enables Sv57 support.
Timmmm
force-pushed
the
user/timh/ptw_nice_types
branch
from
4063c08 to
da2538a
Compare
| // and does not legalize the asid field if asidlen is not the maximum. | ||
| function legalize_satp( | ||
| a : Architecture, | ||
| o : xlenbits, |
There was a problem hiding this comment.
Suggest adding comment: /* previous value of satp */
| function legalize_satp( | ||
| a : Architecture, | ||
| o : xlenbits, | ||
| v : xlenbits, |
There was a problem hiding this comment.
Suggest adding comment: /* proposed new value of satp */
|
|
||
| // PUBLIC: invoked in init_vmem() [riscv_vmem.sail] | ||
| function init_TLB() -> unit = { |
There was a problem hiding this comment.
Will we not need 'init_TLB()' if we allow a dynamic reset on the Sail model?
| @@ -127,7 +127,6 @@ function loop () : unit -> unit = { | |||
| function init_model () -> unit = { | |||
| init_platform (); /* devices */ | |||
| init_sys (); /* processor */ | |||
| init_vmem (); /* virtual memory */ | |||
There was a problem hiding this comment.
Will we not need 'init_vmem()' if we allow a dynamic reset on the Sail model?
There was a problem hiding this comment.
I have reviewed all 15 changed files; all generally LGTM.
Some very minor suggestions (see individual comments on particular files/lines).
I did not study the translation algorithm in riscv_vmem.sail in detail; it looked similar (structure, flow) to original, just fixing it up for the new tighter vmem types, so I think it's probably fine.
The previous version was less type-safe, using 64 bits everywhere, but could work for dynamic choice between Sv32, Sv39, Sv48 and Sv57. I think perhaps this new code requires this to be a static choice, to enable type-checking?
Change the VM code from using
bits(64)everywhere to using more accurate types. This also enables Sv57 support.