closures sometimes boxing closed variables

[adamkapor]

I ran into a case where a closure that I expected to be fast (closed variables of known type, read-only) was not. I think I reduced it to the example below. Should the closure returned by f1 have inferred the type of y?

function f1(x)
    y=x
    return z->y*z
end

function f2(x)
    return z->x*z
end

g1 = f1(1.)
g1.y  #Box(1.0)

g2 = f2(1.)
g2.x  #1.0

This is with a recent version:

Julia Version 0.5.0-dev+3149
Commit 3c72273 (2016-03-14 16:15 UTC)
Platform Info:
  System: Darwin (x86_64-apple-darwin15.2.0)
  CPU: Intel(R) Core(TM) i7-4870HQ CPU @ 2.50GHz
  WORD_SIZE: 64
  BLAS: libopenblas (USE64BITINT DYNAMIC_ARCH NO_AFFINITY Haswell)
  LAPACK: libopenblas64_
  LIBM: libopenlibm
  LLVM: libLLVM-3.7.1 (ORCJIT, haswell)

@code_llvm shows that g2 is nice but g1 is not so nice:

julia> @code_llvm g1(1.)

define %jl_value_t* @"julia_#1_23995"(%jl_value_t*, double) #0 {
top:
  %2 = alloca [9 x %jl_value_t*], align 8
  %.sub = getelementptr inbounds [9 x %jl_value_t*], [9 x %jl_value_t*]* %2, i64 0, i64 0
  %3 = getelementptr [9 x %jl_value_t*], [9 x %jl_value_t*]* %2, i64 0, i64 4
  %4 = getelementptr [9 x %jl_value_t*], [9 x %jl_value_t*]* %2, i64 0, i64 7
  %5 = getelementptr [9 x %jl_value_t*], [9 x %jl_value_t*]* %2, i64 0, i64 2
  store %jl_value_t* null, %jl_value_t** %5, align 8
  %6 = getelementptr [9 x %jl_value_t*], [9 x %jl_value_t*]* %2, i64 0, i64 3
  store %jl_value_t* null, %jl_value_t** %6, align 8
  store %jl_value_t* null, %jl_value_t** %3, align 8
  %7 = getelementptr [9 x %jl_value_t*], [9 x %jl_value_t*]* %2, i64 0, i64 5
  store %jl_value_t* null, %jl_value_t** %7, align 8
  %8 = getelementptr [9 x %jl_value_t*], [9 x %jl_value_t*]* %2, i64 0, i64 6
  store %jl_value_t* null, %jl_value_t** %8, align 8
  store %jl_value_t* null, %jl_value_t** %4, align 8
  %9 = getelementptr [9 x %jl_value_t*], [9 x %jl_value_t*]* %2, i64 0, i64 8
  store %jl_value_t* null, %jl_value_t** %9, align 8
  %10 = bitcast [9 x %jl_value_t*]* %2 to i64*
  store i64 14, i64* %10, align 8
  %11 = load i64, i64* bitcast (%jl_value_t*** @jl_tls_states to i64*), align 8
  %12 = getelementptr [9 x %jl_value_t*], [9 x %jl_value_t*]* %2, i64 0, i64 1
  %13 = bitcast %jl_value_t** %12 to i64*
  store i64 %11, i64* %13, align 8
  store %jl_value_t** %.sub, %jl_value_t*** @jl_tls_states, align 8
  store %jl_value_t* %0, %jl_value_t** %5, align 8
  %14 = bitcast %jl_value_t* %0 to i64*
  %15 = load i64, i64* %14, align 8
  %16 = bitcast %jl_value_t** %4 to i64*
  store i64 %15, i64* %16, align 8
  store %jl_value_t* inttoptr (i64 13063272696 to %jl_value_t*), %jl_value_t** %9, align 8
  %17 = call %jl_value_t* @jl_f_getfield(%jl_value_t* null, %jl_value_t** %4, i32 2)
  store %jl_value_t* %17, %jl_value_t** %7, align 8
  store %jl_value_t* inttoptr (i64 4458931072 to %jl_value_t*), %jl_value_t** %3, align 8
  %18 = call %jl_value_t* @jl_gc_alloc_1w()
  %19 = getelementptr inbounds %jl_value_t, %jl_value_t* %18, i64 -1, i32 0
  store %jl_value_t* inttoptr (i64 4456492784 to %jl_value_t*), %jl_value_t** %19, align 8
  %20 = bitcast %jl_value_t* %18 to double*
  store double %1, double* %20, align 8
  store %jl_value_t* %18, %jl_value_t** %8, align 8
  %21 = call %jl_value_t* @jl_apply_generic(%jl_value_t** %3, i32 3)
  store %jl_value_t* %21, %jl_value_t** %6, align 8
  %22 = load i64, i64* %13, align 8
  store i64 %22, i64* bitcast (%jl_value_t*** @jl_tls_states to i64*), align 8
  ret %jl_value_t* %21
}

julia> @code_llvm g2(1.)

define double @"julia_#3_23996"(%"##3#4"*, double) #0 {
top:
  %2 = getelementptr inbounds %"##3#4", %"##3#4"* %0, i64 0, i32 0
  %3 = load double, double* %2, align 8
  %4 = fmul double %3, %1
  ret double %4
}

[KristofferC]

See: #15276, #15358

[yuyichao]

Close as dup.