Regression in pow_fast

[pitx-perf]

Arising from #54512.

MWE:

function norm_me_base(V)
    return sqrt(V[1]^2 + V[2]^2 + V[3]^2)
end

@fastmath function norm_me_fast(V)
    return sqrt(V[1]^2 + V[2]^2 + V[3]^2)
end

function norm_me_base_ntimes(V, n::Int)
    y = 0.0
    for i = 1:n
        y += norm_me_base(V)
    end
    return y
end

function norm_me_fast_ntimes(V, n::Int)
    y = 0.0
    for i = 1:n
        y += norm_me_fast(V)
    end
    return y
end

using BenchmarkTools
using StaticArrays

function benchmark_that()
v = SVector{3}(10.567, 5.4, 2.28)

@btime norm_me_base(v)
# 1.11   2.800 ns (0 allocations: 0 bytes)
# 1.12   2.800 ns (0 allocations: 0 bytes)

@btime norm_me_fast(v)
# 1.11   2.900 ns (0 allocations: 0 bytes)
# 1.12   82.102 ns (0 allocations: 0 bytes)

 @btime norm_me_base_ntimes(v, 1000)
# 1.11   850.667 ns (0 allocations: 0 bytes)
# 1.12   800.000 ns (0 allocations: 0 bytes)

@btime norm_me_fast_ntimes(v, 1000)
# 1.11   857.143 ns (0 allocations: 0 bytes)
# 1.12    79.100 μs (0 allocations: 0 bytes)
end

Function	Julia 1.11	Julia 1.12
norm_me_base(v)	2.800 ns (0 allocations: 0 bytes)	2.800 ns (0 allocations: 0 bytes)
norm_me_fast(v)	2.900 ns (0 allocations: 0 bytes)	82.102 ns (0 allocations: 0 bytes)
norm_me_base_ntimes(v, 1000)	850.667 ns (0 allocations: 0 bytes)	800.000 ns (0 allocations: 0 bytes)
norm_me_fast_ntimes(v, 1000)	857.143 ns (0 allocations: 0 bytes)	79.100 μs (0 allocations: 0 bytes)

[oscardssmith]

That doesn't make much sense. This should be going through literal_pow...

Edit: Looks like the @fastmath macro breaks literal_pow (ironically for "fast" math).

[KristofferC]

That doesn't make much sense. This should be going through literal_pow...
Edit: Looks like the @fastmath macro breaks literal_pow (ironically for "fast" math).

LLVM can optimize the powi intrinsic for constant powers:

julia> my_pow_fast(x::Float64, y::Integer) = ccall("llvm.powi.f64.i32", llvmcall, Float64, (Float64, Int32), x, y)
my_pow_fast (generic function with 1 method)

julia> f(a::Float64) = my_pow_fast(a, 2)
f (generic function with 2 methods)

julia> @code_llvm f(3.0)
; Function Signature: f(Float64)
;  @ REPL[30]:1 within `f`
define double @julia_f_8765(double %"a::Float64") #0 {
top:
; ┌ @ REPL[29]:1 within `my_pow_fast`
   %0 = fmul double %"a::Float64", %"a::Float64"
   ret double %0
; └
}

However, after #54512 the pow_fast function is no longer inlined, so you get a normal function call.

#59824 should fix that