DARWIN/ARM64: Bad codegen for sign extend

[Keno]

julia> function f(n)
       (n_positive, neg) = Base.split_sign(n); ccall(:jl_breakpoint, Cvoid, ())
       Base.dec(n_positive, 1, neg)
       end

LLVM IR:

;  @ REPL[9]:2 within `f'
; ┌ @ intfuncs.jl:743 within `split_sign'
; │┌ @ int.jl:172 within `abs'
; ││┌ @ int.jl:132 within `flipsign'
     %7 = ashr i32 %0, 31
     %8 = add i32 %0, %7
     %9 = xor i32 %8, %7
; │└└
; │┌ @ promotion.jl:360 within `<'
; ││┌ @ promotion.jl:291 within `promote'
; │││┌ @ promotion.jl:268 within `_promote'
; ││││┌ @ number.jl:7 within `convert'
; │││││┌ @ boot.jl:757 within `Int64'
; ││││││┌ @ boot.jl:676 within `toInt64'
         %10 = sext i32 %0 to i64

Assembly:

; │┌ @ intfuncs.jl:743 within `split_sign'
; ││┌ @ int.jl:172 within `abs'
; │││┌ @ int.jl:132 within `flipsign'
	asr	w10, w9, #31
	add	w11, w9, w10
	eor	w0, w11, w10
; ││└└
; ││┌ @ promotion.jl:360 within `<'
; │││┌ @ promotion.jl:291 within `promote'
; ││││┌ @ promotion.jl:268 within `_promote'
; │││││┌ @ number.jl:7 within `convert'
; ││││││┌ @ boot.jl:757 within `Int64'
; │││││││┌ @ boot.jl:676 within `toInt64'
	mov	x8, x9
; │││└└└└└
; │││ @ promotion.jl:360 within `<' @ int.jl:83
	cmp	x8, #0                          ; =0
	cset	w10, lt

w9 is not sign extended in the x9 register

[Keno]

I guess another piece I should have mentioned is that it's spilling the register:

	str	w0, [sp, #8]                    ; 4-byte Folded Spill
	blr	x8
	ldr	w9, [sp, #8]                    ; 4-byte Folded Reload

If it had used the original x0, there wouldn't be an issue because the value comes in sign extended.

[Keno]

Test case: https://gist.github.com/Keno/99c49ee7f3e4f262e5ae8551dbc585fa

[aemerson]

I commented on the LLVM bug report, but wanted to ask here as well: what's stopping adoption of the default instruction selector at -O0 (GlobalISel) over using FastISel?

[StefanKarpinski]

My guess would be speed, but perhaps @Keno has a more specific reason.

[Keno]

No, there isn't any good reason other than that nobody has tried it and measured the impact.

[aemerson]

In that case I'd recommend that Julia just use the default selector, since GlobalISel is actively maintained.