[AArch64] Sub-optimal folding of a != -a

[SamTebbs33]

The C++ code:

bool foo (unsigned a)
{
  return a != -a;
}

on AArch64 LLVM -O2 generates:

foo(unsigned int):                                // @foo(unsigned int)
        lsl     w8, w0, #1
        cmp     w8, #0
        cset    w0, ne
        ret

GCC does the more compact:

foo(unsigned int):
        cmn     w0, w0
        cset    w0, ne
        ret

Godbolt link: https://godbolt.org/z/z8Mnhh6q4

[llvmbot]

@llvm/issue-subscribers-backend-aarch64

[vfdff]

https://reviews.llvm.org/D144324

[vfdff]

new solution https://reviews.llvm.org/D145846, and its assemble will similar as following:

foo(unsigned int):
	tst	w0, #0x7fffffff
	cset	w0, ne
	ret