我编写了一个小函数来获取结构内部字段的字节偏移，但是要实现这一点，我每次都要在字段上进行迭代，直到找到符号为止。但是在C语言中，offsetof()仅在编译时计算一次，因为偏移量不再在结构内部改变。我想知道如何尽可能地优化它，因为它是一个非常基本的功能，我需要经常为每个帧调用此功能（3D游戏引擎）。
function offsetof(type_, member::Symbol)
for (i, item) in enumerate(fieldnames(type_))
if item == member
return fieldoffset(type_, i)
end
#print(typeof(i))
end
# what to do when symbol not in type_?
throw("$type_ has no member named $member")
end
用法：
type ABC
a::Int64
b::Int64
c::Int64
end
offsetof(ABC, :a) # 0
offsetof(ABC, :b) # 8
offsetof(ABC, :c) # 16
查看LLVM代码，其中有很多代码：
@code_llvm offsetof(Model_s, :fov)
输出：
; Function Attrs: uwtable
define i64 @julia_offsetof_61750(i8**, i8**) #0 !dbg !5 {
top:
%2 = call i8**** @jl_get_ptls_states() #6
%3 = alloca [13 x i8**], align 8
%.sub = getelementptr inbounds [13 x i8**], [13 x i8**]* %3, i64 0, i64 0
%4 = getelementptr [13 x i8**], [13 x i8**]* %3, i64 0, i64 8
%5 = getelementptr [13 x i8**], [13 x i8**]* %3, i64 0, i64 2
%6 = getelementptr [13 x i8**], [13 x i8**]* %3, i64 0, i64 3
%7 = getelementptr [13 x i8**], [13 x i8**]* %3, i64 0, i64 4
%8 = getelementptr [13 x i8**], [13 x i8**]* %3, i64 0, i64 5
%9 = getelementptr [13 x i8**], [13 x i8**]* %3, i64 0, i64 6
%10 = getelementptr [13 x i8**], [13 x i8**]* %3, i64 0, i64 7
%11 = bitcast i8*** %4 to i8*
call void @llvm.memset.p0i8.i32(i8* %11, i8 0, i32 40, i32 8, i1 false)
%12 = bitcast [13 x i8**]* %3 to i64*
%13 = bitcast i8*** %5 to i8*
call void @llvm.memset.p0i8.i64(i8* %13, i8 0, i64 40, i32 8, i1 false)
store i64 22, i64* %12, align 8
%14 = getelementptr [13 x i8**], [13 x i8**]* %3, i64 0, i64 1
%15 = bitcast i8**** %2 to i64*
%16 = load i64, i64* %15, align 8
%17 = bitcast i8*** %14 to i64*
store i64 %16, i64* %17, align 8
store i8*** %.sub, i8**** %2, align 8
store i8** null, i8*** %10, align 8
%18 = getelementptr [13 x i8**], [13 x i8**]* %3, i64 0, i64 12
%19 = getelementptr [13 x i8**], [13 x i8**]* %3, i64 0, i64 11
%20 = getelementptr [13 x i8**], [13 x i8**]* %3, i64 0, i64 10
%21 = getelementptr [13 x i8**], [13 x i8**]* %3, i64 0, i64 9
%22 = call i8** @jlsys_fieldnames_43495(i8** inttoptr (i64 375726352 to i8**))
store i8** %22, i8*** %5, align 8
%23 = getelementptr inbounds i8*, i8** %22, i64 1
%24 = bitcast i8** %23 to i64*
%25 = getelementptr i8*, i8** %22, i64 3
%26 = bitcast i8** %25 to i64*
%27 = bitcast i8** %22 to i8****
%28 = bitcast i8**** %2 to i8*
br label %L3
L3: ; preds = %cont2, %top
%"#temp#1.sroa.4.0" = phi i64 [ 1, %top ], [ %47, %cont2 ]
%29 = load i64, i64* %24, align 8
%30 = add i64 %29, 1
%31 = icmp eq i64 %"#temp#1.sroa.4.0", %30
br i1 %31, label %L33, label %if
if: ; preds = %L3
%32 = add i64 %"#temp#1.sroa.4.0", -1
%33 = load i64, i64* %26, align 8
%34 = icmp ult i64 %32, %33
br i1 %34, label %idxend, label %oob
L33: ; preds = %L3
store i8** inttoptr (i64 77348880 to i8**), i8*** %4, align 8
store i8** inttoptr (i64 77362488 to i8**), i8*** %21, align 8
store i8** inttoptr (i64 375726352 to i8**), i8*** %20, align 8
store i8** inttoptr (i64 377039808 to i8**), i8*** %19, align 8
store i8** %1, i8*** %18, align 8
%35 = call i8** @"jsys1_#print_to_string#229_39171"(i8** inttoptr (i64 77362448 to i8**), i8*** %4, i32 5)
store i8** %35, i8*** %10, align 8
call void @jl_throw(i8** %35)
unreachable
oob: ; preds = %if
%36 = alloca i64, align 8
store i64 %"#temp#1.sroa.4.0", i64* %36, align 8
call void @jl_bounds_error_ints(i8** %22, i64* nonnull %36, i64 1)
unreachable
idxend: ; preds = %if
%37 = load i8***, i8**** %27, align 8
%38 = getelementptr i8**, i8*** %37, i64 %32
%39 = load i8**, i8*** %38, align 8
%40 = icmp eq i8** %39, null
br i1 %40, label %fail, label %cont2
fail: ; preds = %idxend
call void @jl_throw(i8** inttoptr (i64 84505768 to i8**))
unreachable
cont2: ; preds = %idxend
store i8** %39, i8*** %6, align 8
%41 = call i8** @jl_gc_pool_alloc(i8* %28, i32 1512, i32 32)
%42 = getelementptr i8*, i8** %41, i64 -1
%43 = bitcast i8** %42 to i8***
store i8** inttoptr (i64 109041296 to i8**), i8*** %43, align 8
store i8** %41, i8*** %7, align 8
%44 = getelementptr i8*, i8** %41, i64 1
%45 = bitcast i8** %44 to i8***
%46 = bitcast i8** %41 to i64*
store i64 %"#temp#1.sroa.4.0", i64* %46, align 16
store i8** %39, i8*** %45, align 8
%47 = add i64 %"#temp#1.sroa.4.0", 1
store i8** %39, i8*** %8, align 8
store i8** %39, i8*** %9, align 8
%48 = icmp eq i8** %39, %1
br i1 %48, label %if3, label %L3
if3: ; preds = %cont2
%sext = shl i64 %"#temp#1.sroa.4.0", 32
%49 = ashr exact i64 %sext, 32
%50 = icmp eq i64 %49, %"#temp#1.sroa.4.0"
br i1 %50, label %pass5, label %fail4
fail4: ; preds = %if3
call void @jl_throw(i8** inttoptr (i64 77358144 to i8**))
unreachable
pass5: ; preds = %if3
%51 = trunc i64 %"#temp#1.sroa.4.0" to i32
%52 = call i64 inttoptr (i64 1693737504 to i64 (i8**, i32)*)(i8** inttoptr (i64 375726352 to i8**), i32 %51)
%53 = load i64, i64* %17, align 8
store i64 %53, i64* %15, align 8
ret i64 %52
}
最佳答案
这可能是Base.@pure表示法的用例。对于问题中的示例，它将去：
# note the Base.@pure notation at beginning of definition:
Base.@pure function offsetof(type_, member::Symbol)
for (i, item) in enumerate(fieldnames(type_))
if item == member
return fieldoffset(type_, i)
end
#print(typeof(i))
end
# what to do when symbol not in type_?
throw("$type_ has no member named $member")
end
现在，我们可以：
type ABC
a::Int64
b::Int64
c::Int64
end
offsetof(ABC, :a) # 0
并且
@code_llvm的offsetof仍然很长。但是，如果我们在函数中使用它，Julia将在编译时运行offsetof（这是允许的，因为纯函数应该依赖于它们的参数，除了返回从它们计算出的值外，什么也不做）。例如：julia> f() = offsetof(ABC,:b)
f (generic function with 1 method)
julia> f()
0x0000000000000008
julia> @code_llvm f()
define i64 @julia_f_63287() #0 !dbg !5 {
top:
%ptls_i8 = call i8* asm "movq %fs:0, $0;\0Aaddq $$-10928, $0", "=r,~{dirflag},~{fpsr},~{flags}"() #1
ret i64 8
}
请注意，
f()仅由ret i64 8组成，该0x8返回offsetof(ABC,:b)，即的值。这是描述@pure表示法的链接：https://github.com/JuliaLang/julia/issues/414（另请参见：https://github.com/JuliaLang/julia/issues/14324）
请注意：纯符号可能会不断变化，并且可能在版本之间变化。它在0.6中起作用。将来，我们可以使用更智能的编译器（不断进行传播）来更轻松地实现此优化。
关于optimization - 将offsetof()优化到恒定值？，我们在Stack Overflow上找到一个类似的问题：https://stackoverflow.com/questions/44668696/