Re: [patch][rfa] CGEN Model Support for SID

No comments on this? Too much or too little information?

OK to commit?

Dave

Dave Brolley wrote:

> Hi,
>
> This submission cleans up the partially adapted support for CGEN 
> models for SID and enables it. Attached are patches for CGEN and SID 
> which accomplish this:
>
> o sid-model.patch.txt's: Patches for CGEN and SID
> o cgen-model.h: New class definition in sid/component/cgen-cpu
> o *.ChangeLog: ChangeLogs for the above
>
> Also attached are examples of the generated code and suggested usage 
> for the xstormy16:
>
> o xstormy16-model.{cxx,h}: CGEN-generated model code
> o xstormy16-model-gen.patch.txt: Additional code generated into 
> xstormy16-decode.h and xstormy16-sem.cxx
> o xstormy16-model.patch.txt: Patch showing suggest usage of the 
> generated code by a SID cgen-cpu.
>
> Notes:
>
> o The implementation is similar to the existing implementation for 
> sim/common simulators except:
>  - The generated functions are encapsulated into a class: 
> <prefix>_<model>_model
>  - Functions are generated for modelling each insn both before and 
> after execution. The sim/common implementation only generated a 
> function to be called after execution, however, there was a need for 
> modelling both before and after in the implementaion of the FRV 
> simulator which was handled somewhat awkwardly. Therefore I added this 
> capability for this implementation.
>  - As in the existing sim/common implementation, there are hooks 
> provided (at the cgen-model level) for modelling  before and after 
> groups of parallel insns.
>
> A general outline of the design follows:
>
> o A new class, cgen-model, has been defined in 
> sid/component/cgen-cpu/cgen-model.h. It provides the methods for 
> modelling before and after groups of parallel insns with default 
> virtual implementations and keeps a pointer to the cpu for reporting 
> computed latencies.
>
> o The script sid/component/cgen-cpu/CGEN.sh has been enhanced to 
> process the option 'with-profile=fn' and to generate the files 
> model.cxx and model.h.
>
> o For each model (cgen model) used by a all machines (cgen mach) being 
> generated, a class, @prefix@_@model@_model is generated in 
> sid/component/cgen-cpu/@cpu@/@prefix@-model.{cxx,h}. These classes 
> each inherit from cgen-cpu and implement the functions for modelling 
> each unit before and after each insn. Default virtual implementations 
> are provided as well as a table mapping each insn to its unit 
> modelling functions and functions for indexing the table to make the 
> calls.
>
> o The generated code, as is, is complete and implements default 
> latency computations as defined by the (timing ...) elements of each 
> insn in the .cpu file.
>
> o Since the default implementations are virtual, the programmer may 
> inherit the generated class(es) to implement more complex modelling 
> behavior for each unit of each model as needed.
>
> o Calls to the modelling functions are inserted, as necessary, into 
> the cpu's step_insns method. See xstormy16-model.patch.txt for an 
> example.
>
> This implementation has been tested on xstormy16 and an internal port 
> which makes extensive use of this new work. Other ports are not 
> affected since they do not make use of this functionality.
>
> I am seeking comments leading toward approval to commit the general 
> changes. I can also commit the xtormy16-specific code if folks think 
> it would be of use as an example.
>
> Dave