rsqrt as builtin

[expikr]

for the hardware instructions where available

[nektro]

do you have docs on what this function is?

[nektro]

1 / @sqrt(x) should compile to what you seek

[expikr]

does it always, for vectors as well?

Code

export fn rsqrt(num: f32) f32 {
    return 1/@sqrt(num);
}

Godbolt default

# Compilation provided by Compiler Explorer at https://godbolt.org/
.LCPI0_0:
        .long   0x3f800000
rsqrt:
        push    rbp
        mov     rbp, rsp
        sub     rsp, 4
        vmovaps xmm1, xmm0
        vmovss  dword ptr [rbp - 4], xmm1
        vsqrtss xmm1, xmm0, xmm1
        vmovss  xmm0, dword ptr [rip + .LCPI0_0]
        vdivss  xmm0, xmm0, xmm1
        add     rsp, 4
        pop     rbp
        ret

Godbolt -O ReleaseFast

# Compilation provided by Compiler Explorer at https://godbolt.org/
.LCPI0_0:
        .long   0x3f800000
rsqrt:
        vsqrtss xmm0, xmm0, xmm0
        vmovss  xmm1, dword ptr [rip + .LCPI0_0]
        vdivss  xmm0, xmm1, xmm0
        ret

[nektro]

at a glance it looks like it should be using (V)RSQRTPS instead

[castholm]

Somewhat related: #767
At a quick glance there doesn't appear to be an LLVM intrinsic for rsqrt, but I don't think that's necessarily an argument against adding @rsqrt and I can such a builtin being useful in certain cases.

1 / @sqrt(x) should compile to what you seek

Unless @setFloatMode(.optimized) or equivalent compiler flags are in effect I would hope that 1 / @sqrt(x) is never compiled to reciprocal square root, for the same reasons it is undesired for a * b + c to silently compile to fused multiply-adds, divisions to reciprocals, etc. that change the semantics of the expression.

[expikr]

Continue tracking in #767 (comment)