Support setting MCPU during Make, for ARM

[IanButterworth]

As is outlined in https://community.arm.com/developer/tools-software/tools/b/tools-software-ides-blog/posts/compiler-flags-across-architectures-march-mtune-and-mcpu

In short, you cannot support both x86 and Arm with exactly the same compiler flags. You must use -mcpu for Arm and -march for x86.

I also opened up JuliaCI/julia-buildbot#149 to switch arm builds to mcpu, before realizing I needed to also adjust the Make file here

[IanButterworth]

There might be a better way to do the OR logic here

[vtjnash]

Suggested change

	ifdef do_step
	ifneq ($(MARCH)$(MCPU),)

[IanButterworth]

Great, thanks. Incorporated

[yuyichao]

I think you can just pass in CFLAGS etc? Different from MARCH it's not like this is going to be used anywhere else.

Also, this isn't very ARM specific (and FWIW the condition doesn't seem to capture aarch64 which is arguably more important) though I think it could make sense to forward native march to mtune automatically. According to that article this should work on both x86 and arm.

[IanButterworth]

The paradigm just seems to be different for ARM, so I thought it best to directly support MCPU?

Also, that article claims

LLVM compilers only support “native” for the -mcpu and -mtune flags. You cannot use -march=native with LLVM-based compilers.

Is that right? If march=native wasn't being passed to LLVM for user code because of that, perhaps there's reason to now add -mcpu=native for user code?

I'm wading into an area I'm not familiar with so apologies if these are flawed suggestions

[IanButterworth]

And on the condition for catching arm & aarch64, you're right.. I need to think of a better catch there

[yuyichao]

Is that right?

Not sure if that's right but this PR doesn't change that it seems.

If march=native wasn't being passed to LLVM for user code because of that, perhaps there's reason to now add -mcpu=native for user code?

All of the flags mentioned here (march, mcpu, mtune), no matter what you passed in, has nothing to do with user code. MARCH is used as default for JULIA_CPU_TARGET but even that has zero effect on the user code.

[IanButterworth]

Not sure if that's right but this PR doesn't change that it seems.

Sorry, I didn't mean to imply it was. That was a tangential question. I think I'm concluding that I don't think I know enough about the internals of the language to venture any further.

I just tweaked the logic to catch *arm* and *aarch64* in the XC_HOST section. Maybe there's a more robust method

[yuyichao]

Sorry, I didn't mean to imply it was. That was a tangential question. I think I'm concluding that I don't think I know enough about the internals of the language to venture any further.

What I meant is that if it is currently used and we didn't get an error then adding -mtune should be enough...

[IanButterworth]

Edit: Done
As a reminder to myself, if this goes ahead we'd probably want to update the advice in the Arm build tips documentation to use MCPU=native

[vtjnash]

What does filter do without a filter list? Is this the same as strip?

[IanButterworth]

On review I think this wouldn't have worked. I've updated to the OR format you suggested for the other case. Thanks