Large performance problem due to @fastmath on many operations

[ChrisRackauckas]

function f1(a,b,c,d,e,f,g,h,j,k,l,m,n,o,p)
  aidx = eachindex(a)
  for i in aidx
    @inbounds a[i] = b[i]+c*(d*e[i]+f*g[i]+h*j[i]+k*l[i]+m*n[i]+o*p[i])
  end
end
function f2(a,b,c,d,e,f,g,h,j,k,l,m,n,o,p)
  aidx = eachindex(a)
  @fastmath for i in aidx
    @inbounds a[i] = b[i]+c*(d*e[i]+f*g[i]+h*j[i]+k*l[i]+m*n[i]+o*p[i])
  end
end
a = rand(10)
b = rand(10)
c = 0.1
d = 0.1
e = rand(10)
f = 0.1
g = rand(10)
h = 0.1
j = rand(10)
k = 0.1
l = rand(10)
m = 0.1
n = rand(10)
o = 0.1
p = rand(10)

using BenchmarkTools
@benchmark f1($a,$b,$c,$d,$e,$f,$g,$h,$j,$k,$l,$m,$n,$o,$p)
@benchmark f2($a,$b,$c,$d,$e,$f,$g,$h,$j,$k,$l,$m,$n,$o,$p)

@benchmark f1($a,$b,$c,$d,$e,$f,$g,$h,$j,$k,$l,$m,$n,$o,$p)
BenchmarkTools.Trial: 
  memory estimate:  112 bytes
  allocs estimate:  7
  --------------
  minimum time:     83.086 ns (0.00% GC)
  median time:      164.388 ns (0.00% GC)
  mean time:        152.032 ns (5.06% GC)
  maximum time:     2.518 μs (92.10% GC)
  --------------
  samples:          10000
  evals/sample:     983

@benchmark f2($a,$b,$c,$d,$e,$f,$g,$h,$j,$k,$l,$m,$n,$o,$p)
BenchmarkTools.Trial: 
  memory estimate:  5.27 KiB
  allocs estimate:  297
  --------------
  minimum time:     11.124 μs (0.00% GC)
  median time:      22.248 μs (0.00% GC)
  mean time:        20.982 μs (2.20% GC)
  maximum time:     2.470 ms (96.68% GC)
  --------------
  samples:          10000
  evals/sample:     1

Found in the same code as #22255

[yuyichao]

Is this actually a regression?

[ChrisRackauckas]

Oops, my bad. I noticed this before on v0.5 but never reported it. Changed the title. These tests are run on v0.6-rc2.

[ChrisRackauckas]

Just tested the lastest nightly binaries and it looks like it's still the case:

julia> @time f1(a,b,c,d,e,f,g,h,j,k,l,m,n,o,p)
  0.014693 seconds (4.06 k allocations: 225.578 KiB)

julia> @time f2(a,b,c,d,e,f,g,h,j,k,l,m,n,o,p)
  0.024622 seconds (10.12 k allocations: 568.695 KiB)

julia> @time f1(a,b,c,d,e,f,g,h,j,k,l,m,n,o,p)
  0.000004 seconds (4 allocations: 160 bytes)

julia> @time f2(a,b,c,d,e,f,g,h,j,k,l,m,n,o,p)
  0.000043 seconds (294 allocations: 5.313 KiB)

julia> versioninfo()
Julia Version 0.7.0-DEV.491
Commit 45c15682ff* (2017-06-07 05:19 UTC)
Platform Info:
  OS: Windows (x86_64-w64-mingw32)
  CPU: Intel(R) Core(TM) i7-4770K CPU @ 3.50GHz
  WORD_SIZE: 64
  BLAS: libopenblas (USE64BITINT DYNAMIC_ARCH NO_AFFINITY Haswell)
  LAPACK: libopenblas64_
  LIBM: libopenlibm
  LLVM: libLLVM-3.9.1 (ORCJIT, haswell)
Environment:

[ChrisRackauckas]

This is number of operations dependent:

function f1(a,b,c,d,e,f,g,h,j,k,l,m,n)
  aidx = eachindex(a)
  for i in aidx
    @inbounds a[i] = b[i]+c*(d*e[i]+f*g[i]+h*j[i]+k*l[i]+m*n[i])
  end
end
function f2(a,b,c,d,e,f,g,h,j,k,l,m,n)
  aidx = eachindex(a)
  @fastmath for i in aidx
    @inbounds a[i] = b[i]+c*(d*e[i]+f*g[i]+h*j[i]+k*l[i]+m*n[i])
  end
end
a = rand(10)
b = rand(10)
c = 0.1
d = 0.1
e = rand(10)
f = 0.1
g = rand(10)
h = 0.1
j = rand(10)
k = 0.1
l = rand(10)
m = 0.1
n = rand(10)

julia> @benchmark f1($a,$b,$c,$d,$e,$f,$g,$h,$j,$k,$l,$m,$n)
BenchmarkTools.Trial:
  memory estimate:  0 bytes
  allocs estimate:  0
  --------------
  minimum time:     35.421 ns (0.00% GC)
  median time:      36.593 ns (0.00% GC)
  mean time:        39.635 ns (0.00% GC)
  maximum time:     299.767 ns (0.00% GC)
  --------------
  samples:          10000
  evals/sample:     1000

julia> @benchmark f2($a,$b,$c,$d,$e,$f,$g,$h,$j,$k,$l,$m,$n)
BenchmarkTools.Trial:
  memory estimate:  0 bytes
  allocs estimate:  0
  --------------
  minimum time:     34.836 ns (0.00% GC)
  median time:      36.300 ns (0.00% GC)
  mean time:        39.278 ns (0.00% GC)
  maximum time:     129.684 ns (0.00% GC)
  --------------
  samples:          10000
  evals/sample:     1000

But this issue happens even though the parenthesis encloses <16 values?

[fredrikekre]

The examples in this issue seem to be performing much better on master, example in OP:

julia> @benchmark f1($a,$b,$c,$d,$e,$f,$g,$h,$j,$k,$l,$m,$n,$o,$p)
BenchmarkTools.Trial: 
  memory estimate:  112 bytes
  allocs estimate:  7
  --------------
  minimum time:     69.316 ns (0.00% GC)
  median time:      72.273 ns (0.00% GC)
  mean time:        83.480 ns (7.75% GC)
  maximum time:     39.859 μs (99.71% GC)
  --------------
  samples:          10000
  evals/sample:     964

julia> @benchmark f2($a,$b,$c,$d,$e,$f,$g,$h,$j,$k,$l,$m,$n,$o,$p)
BenchmarkTools.Trial: 
  memory estimate:  112 bytes
  allocs estimate:  7
  --------------
  minimum time:     62.282 ns (0.00% GC)
  median time:      68.275 ns (0.00% GC)
  mean time:        79.657 ns (8.59% GC)
  maximum time:     41.918 μs (99.73% GC)
  --------------
  samples:          10000
  evals/sample:     968

[KristofferC]

Looks fixed indeed.