[AArch64] Generate stp for complex repeating constansts

[llvmbot]

Bugzilla Link	52141
Version	trunk
OS	Windows NT
Reporter	LLVM Bugzilla Contributor
CC	@Arnaud-de-Grandmaison-ARM,@DMG862,@smithp35

Extended Description

For complex repeating constants like:

void
foo (unsigned long long *a)
{
  a[0] = 0x0140c0da0140c0daULL;
}

aarch64 clang generates

foo(unsigned long long*):                               // @​foo(unsigned long long*)
        mov     x8, #​49370
        movk    x8, #​320, lsl #​16
        movk    x8, #​49370, lsl #​32
        movk    x8, #​320, lsl #​48
        str     x8, [x0]
        ret

but for constants where the top half is the same as the bottom half we can construct just one half and use a store pair to duplicate it:

foo(unsigned long long*):
        mov     w1, 49370
        movk    w1, 0x140, lsl 16
        stp     w1, w1, [x0]
        ret

See also: https://godbolt.org/z/n3M4qh6sW

[rajdroid]

Hi, may I give it a try? I don't have much experience in the backend but happy to learn and tackle this problem.

[rajdroid]

Adding @Arnaud-de-Grandmaison-ARM @smithp35 for suggestions.

[kbeyls]

@ilinpv was the original reporter of this issue it seems. @rajdroid I'm assuming he'll know best if anyone else is already planning to work on this. I'd say, don't hold back in giving this a try. If you run into issues or are not sure how to start, please don't hold back in asking questions. Good places to ask questions are here, on this issue; maybe on LLVM's Discord; and on the llvm-dev mailing list.

[rajdroid]

Thanks @kbeyls, I'll start working on this issue then, and I'll ask questions if I run into any issues.

[ilinpv]

As far as I know, currently nobody is working on this. @rajdroid, you are welcome, it would be great if you give this a try.

[rajdroid]

Thanks @ilinpv for letting me know. Sure, I'll give this a try.

I have few basic questions.

1. How can I debug the AArch64 backend on an x86 host?
2. Can I use Qemu with Arm host?

[ilinpv]

1. How can I debug the AArch64 backend on an x86 host?

You can build x86 llvm on host and debug it, specifying AArch64 target, for example "clang -target aarch64-none-linux-gnu"

2. Can I use Qemu with Arm host?

Yes, but I am not sure if it is needed for this issue.

[rajdroid]

Oh I see, I already have x86 llvm debug build, I will use that then. Thanks!

[ilinpv]

@rajdroid if you are still working on this may I ask you to become assignee of the bug ?

[llvmbot]

@llvm/issue-subscribers-good-first-issue

[ParkHanbum]

I am implementing it in AArch64LoadStoreOptimizer.
Is it better to implement it in AArch64MIPeepholeOpt?

[ilinpv]

I am implementing it in AArch64LoadStoreOptimizer. Is it better to implement it in AArch64MIPeepholeOpt?

AArch64MIPeepholeOpt is not dealing with stores, AArch64LoadStoreOptimizer seems better place to me.