isbits is slow?

[vtjnash]

julia> x = [1:100]

julia> @sprofile for i = 1:1e7 copy!(x, x) end

julia> Profile.sprofile_tree()
     819 client.jl; _start; line: 372
      819 client.jl; run_repl; line: 166
         819 client.jl; eval_user_input; line: 91
            769 /Users/jameson/.julia/Profile/src/sprofile.jl; anonymous; line: 1
              578 array.jl; copy!; line: 64
               9   array.jl; copy!; line: 58
               549 array.jl; copy!; line: 61
                502 array.jl; unsafe_copy!; line: 47
                 470 reflection.jl; isbits; line: 36
                32  array.jl; unsafe_copy!; line: 48
                 15 array.jl; unsafe_copy!; line: 41
                 4  array.jl; unsafe_copy!; line: 43
                11  array.jl; unsafe_copy!; line: 54
                2   reflection.jl; isbits; line: 36
            43  /Users/jameson/.julia/Profile/src/sprofile.jl; anonymous; line: 309

julia> sprofile_clear()

Similarly using setindex! notation, although the creation of the colon object is about equally expensive.

julia> @sprofile for i = 1:1e7 x[:] = x end

julia> Profile.sprofile_tree()
     1564 client.jl; _start; line: 372
      1564 client.jl; run_repl; line: 166
         1564 client.jl; eval_user_input; line: 91
            1524 /Users/jameson/.julia/Profile/src/sprofile.jl; anonymous; line: 1
              1    array.jl; setindex!; line: 433
              1    range.jl; colon; line: 38
               5   abstractarray.jl; endof; line: 25
               14  array.jl; setindex!; line: 431
               566 array.jl; setindex!; line: 432
                9   array.jl; copy!; line: 58
                556 array.jl; copy!; line: 61
                 501 array.jl; unsafe_copy!; line: 47
                  460 reflection.jl; isbits; line: 36
                 40  array.jl; unsafe_copy!; line: 48
                  18 array.jl; unsafe_copy!; line: 41
                  10 array.jl; unsafe_copy!; line: 43
                 7   array.jl; unsafe_copy!; line: 54
                 1   reflection.jl; isbits; line: 36
               9   array.jl; setindex!; line: 433
               4   promotion.jl; max; line: 189
               1   range.jl; Range1; line: 28
               43  range.jl; colon; line: 38
                10 promotion.jl; max; line: 189
                6  range.jl; Range1; line: 27
                7  range.jl; Range1; line: 28
            21   /Users/jameson/.julia/Profile/src/sprofile.jl; anonymous; line: 309

[vtjnash]

It appears to now be inline, but still is allocating a new box for the result.

(edit: code added)

julia> f() = isbits(Int)
# methods for generic function f
f() at none:1

julia> finfer(f,())
:($(Expr(:lambda, {}, {{},{},{}}, quote  # none, line 1:
        return top(box)(Bool,top(and_int)(top(box)(Bool,top(not_int)(top(getfield)(Int,:mutable)::Bool))::Bool,top(getfield)(Int,:pointerfree)::Bool))::Bool
    end)))

julia> f()
true

julia> disassemble(f,())
WARNING: disassemble(f::Function,t::Tuple) is deprecated, use code_llvm(f,t) instead.

define i1 @julia_f1097() {
top:
  %0 = alloca [4 x %jl_value_t*], align 8
  %.sub = getelementptr inbounds [4 x %jl_value_t*]* %0, i64 0, i64 0
  %1 = getelementptr [4 x %jl_value_t*]* %0, i64 0, i64 2, !dbg !5659
  store %jl_value_t* inttoptr (i64 4 to %jl_value_t*), %jl_value_t** %.sub, align 8
  %2 = load %jl_value_t*** @jl_pgcstack, align 8, !dbg !5659
  %3 = getelementptr [4 x %jl_value_t*]* %0, i64 0, i64 1, !dbg !5659
  %.c = bitcast %jl_value_t** %2 to %jl_value_t*, !dbg !5659
  store %jl_value_t* %.c, %jl_value_t** %3, align 8, !dbg !5659
  store %jl_value_t** %.sub, %jl_value_t*** @jl_pgcstack, align 8, !dbg !5659
  store %jl_value_t* null, %jl_value_t** %1, align 8
  %4 = getelementptr [4 x %jl_value_t*]* %0, i64 0, i64 3
  store %jl_value_t* null, %jl_value_t** %4, align 8
  %5 = load %jl_value_t** inttoptr (i64 4346643568 to %jl_value_t**), align 16, !dbg !5664
  store %jl_value_t* %5, %jl_value_t** %1, align 8, !dbg !5664
  store %jl_value_t* inttoptr (i64 4313910528 to %jl_value_t*), %jl_value_t** %4, align 8, !dbg !5664
  %6 = call %jl_value_t* @jl_f_get_field(%jl_value_t* null, %jl_value_t** %1, i32 2), !dbg !5664
  %7 = getelementptr inbounds %jl_value_t* %6, i64 1, i32 0, !dbg !5664
  %8 = bitcast %jl_value_t** %7 to i8*, !dbg !5664
  %9 = load i8* %8, align 1, !dbg !5664
  %10 = and i8 %9, 1, !dbg !5664
  %11 = icmp eq i8 %10, 0, !dbg !5664
  %12 = load %jl_value_t** inttoptr (i64 4346643568 to %jl_value_t**), align 16, !dbg !5664
  store %jl_value_t* %12, %jl_value_t** %1, align 8, !dbg !5664
  store %jl_value_t* inttoptr (i64 4313910576 to %jl_value_t*), %jl_value_t** %4, align 8, !dbg !5664
  %13 = call %jl_value_t* @jl_f_get_field(%jl_value_t* null, %jl_value_t** %1, i32 2), !dbg !5664
  %14 = getelementptr inbounds %jl_value_t* %13, i64 1, i32 0, !dbg !5664
  %15 = bitcast %jl_value_t** %14 to i8*, !dbg !5664
  %16 = load i8* %15, align 1, !dbg !5664
  %17 = and i8 %16, 1, !dbg !5664
  %18 = icmp ne i8 %17, 0, !dbg !5664
  %19 = and i1 %11, %18, !dbg !5664
  %20 = load %jl_value_t** %3, align 8, !dbg !5664
  %21 = getelementptr inbounds %jl_value_t* %20, i64 0, i32 0, !dbg !5664
  store %jl_value_t** %21, %jl_value_t*** @jl_pgcstack, align 8, !dbg !5664
  ret i1 %19, !dbg !5664
}

[JeffBezanson]

I don't see any allocations there.

[vtjnash]

jl_f_get_field has to allocate a box for the return value (is this a general problem with emit_getfield?)