[PATCH] D71124: [RISCV] support clang driver to select cpu

Mon Apr 6 10:51:24 PDT 2020

khchen added a comment.

@lenary (Sorry for the very late reply...)

In D71124#1829998 <https://reviews.llvm.org/D71124#1829998>, @lenary wrote:

> In D71124#1792216 <https://reviews.llvm.org/D71124#1792216>, @khchen wrote:
>
> > The problem is how `-mcpu` interact with explicitly specified `-march` (or target features).
> >
> > 1. in GCC, -mcpu is only used to chose the pipeline model,
>
>
> I think you mean "in GCC, `-mtune` is only used to choose the pipeline model" (`-mcpu` is not documented in the RISC-V specific GCC options documentation <https://gcc.gnu.org/onlinedocs/gcc/RISC-V-Options.html>).

Yes, It's my mistake, I mean `-mtune`.

> Clang should attempt to maintain compatibility with GCC flags, but if they only implement `-mtune`, then we have a little more freedom to do something ergonomic with `-mcpu`.
> 
> I'll note that clang already has a large TODO around implementing `-mtune` in general, though the AArch64 backend seems to support it for some option choices.
> 
>> 
>> 
>> 2. I also read this <https://community.arm.com/developer/tools-software/tools/b/tools-software-ides-blog/posts/compiler-flags-across-architectures-march-mtune-and-mcpu> article talking about the X86 and ARM to handle those options.
>>   - -march=X: Tells the compiler that X is the minimal architecture the binary must run on.  The compiler is free to use architecture-specific instructions.  This flag behaves differently on Arm and x86.  On Arm, -march does not override -mtune, but on x86 -march will override both -mtune and -mcpu.
>>   - -mtune=X: Tells the compiler to optimize for microarchitecture X but does not allow the compiler to change the ABI or make assumptions about available instructions.  This flag has the more-or-less the same meaning on Arm and x86.
>>   - -mcpu=X: On Arm, this flag is a combination of -march and -mtune.  It simultaneously specifies the target architecture and optimizes for a given microarchitecture. On x86 this flag is a deprecated synonym for -mtune.
>> 
>>     So maybe it makes sense to treat those flags behaves differently on different target .
>> 3. I also tried llc to specific -mcpu and -attr (similar to -march, target-feature) in ARM, -attr will over write the default target-feature in -mcpu.
>> 
>>   on RISC-V, in sometime (or most?) we have same pipeline model but support different extension combination,
> 
> I don't believe this to be correct. lowRISC's Ibex has a completely different pipeline model to rocket, and there are countless other RISC-V cores with different pipeline characteristics, including out-of-order pipeline implementations like BOOM. I don't think we can favour one particular scheduling model (beyond the generic ones we already default to).

Yes, so lowRISC's Ibex need to use different pipleine model.  
I mention it just because in SiFive, the same serial core can share the same schedule model, but they have different isa extension.

>> so I think maybe distinguishing the purpose of -mcpu and -march and make them with no interaction is a good idea. (behavior is equal to GCC)
> 
> In LLVM, if you add `target-cpu` metadata to a function (which is added by clang, based on `-mcpu`), that function will have all the features of that CPU automatically added to it (as if you had used `-mattr` with all the features in the model). If you don't add that metadata, a generic scheduling model will be chosen. This suggests at the moment there can be no separation between `-mtune` and `-march` as there is in GCC (without changes to the target-independent parts of LLVM).

thanks for reminding me, I re-implementation it.

In D71124#1829998 <https://reviews.llvm.org/D71124#1829998>, @lenary wrote:

> In D71124#1792216 <https://reviews.llvm.org/D71124#1792216>, @khchen wrote:
>
> > The problem is how `-mcpu` interact with explicitly specified `-march` (or target features).
> >
> > 1. in GCC, -mcpu is only used to chose the pipeline model,
>
>
> I think you mean "in GCC, `-mtune` is only used to choose the pipeline model" (`-mcpu` is not documented in the RISC-V specific GCC options documentation <https://gcc.gnu.org/onlinedocs/gcc/RISC-V-Options.html>).
>
> Clang should attempt to maintain compatibility with GCC flags, but if they only implement `-mtune`, then we have a little more freedom to do something ergonomic with `-mcpu`.
>
> I'll note that clang already has a large TODO around implementing `-mtune` in general, though the AArch64 backend seems to support it for some option choices.
>
> > 
> > 
> > 2. I also read this <https://community.arm.com/developer/tools-software/tools/b/tools-software-ides-blog/posts/compiler-flags-across-architectures-march-mtune-and-mcpu> article talking about the X86 and ARM to handle those options.
> >   - -march=X: Tells the compiler that X is the minimal architecture the binary must run on.  The compiler is free to use architecture-specific instructions.  This flag behaves differently on Arm and x86.  On Arm, -march does not override -mtune, but on x86 -march will override both -mtune and -mcpu.
> >   - -mtune=X: Tells the compiler to optimize for microarchitecture X but does not allow the compiler to change the ABI or make assumptions about available instructions.  This flag has the more-or-less the same meaning on Arm and x86.
> >   - -mcpu=X: On Arm, this flag is a combination of -march and -mtune.  It simultaneously specifies the target architecture and optimizes for a given microarchitecture. On x86 this flag is a deprecated synonym for -mtune.
> > 
> >     So maybe it makes sense to treat those flags behaves differently on different target .
> > 3. I also tried llc to specific -mcpu and -attr (similar to -march, target-feature) in ARM, -attr will over write the default target-feature in -mcpu.
> > 
> >   on RISC-V, in sometime (or most?) we have same pipeline model but support different extension combination,
>
> I don't believe this to be correct. lowRISC's Ibex has a completely different pipeline model to rocket, and there are countless other RISC-V cores with different pipeline characteristics, including out-of-order pipeline implementations like BOOM. I don't think we can favour one particular scheduling model (beyond the generic ones we already default to).
>
> > so I think maybe distinguishing the purpose of -mcpu and -march and make them with no interaction is a good idea. (behavior is equal to GCC)
>
> In LLVM, if you add `target-cpu` metadata to a function (which is added by clang, based on `-mcpu`), that function will have all the features of that CPU automatically added to it (as if you had used `-mattr` with all the features in the model). If you don't add that metadata, a generic scheduling model will be chosen. This suggests at the moment there can be no separation between `-mtune` and `-march` as there is in GCC (without changes to the target-independent parts of LLVM).

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D71124/new/

https://reviews.llvm.org/D71124