转 h_t_t_p_s_:_/_/quuxplusone.github.io/blog/2021/11/09/pass-string-view-by-value/
Three reasons to pass std::string_view by value
It is idiomatic to pass std::string_view by value. Let’s see why.
First, a little background recap. In C++, everything defaults to pass-by-value; when you say Widget w you actually get a whole new Widget object. But copying big things can be expensive. So we introduce “pass-by-const-reference” as an optimization of “pass-by-value,” and we tell people to pass big and/or expensive things like std::string by const reference instead of by value.
But for small cheap things — int, char*, std::pair, std::span — we continue to prefer the sensible default behavior of pass-by-value.
Pass-by-value has at least three performance benefits over pass-by-(const)-reference. I’ll illustrate all three of them via string_view.
All of these code snippets are presented in isolation, showing either only the callee or only the caller. If the compiler is allowed to see both caller and callee together, and is given -O2, and decides to inline the callee into the caller, then it can usually undo all of the harm caused by an unidiomatic pass-by-reference. So, frequently, you can pass string_view by reference and get away with it. But you should pass string_view by value, so that the compiler doesn’t have to do those heroics on your behalf. And so that your code-reviewer doesn’t have to burn brain cells pondering your unidiomatic decision to pass by reference. Briefly: Pass small cheap types by value! It has only upsides!
Okay, now let’s see those three performance benefits I promised.
1. Eliminate a pointer indirection in the callee
Pass-by-const-reference means that you pass the address of the thing. Pass-by-value means that you pass the thing itself, in registers when possible. (If the thing you’re passing is “non-trivial for purposes of ABI” — such as if it has a non-trivial destructor — then “the thing itself” ends up being passed on the stack, so there will be memory indirection involved either way. But trivial types like int and string_view and span don’t have to worry about that; these types are passed in registers.)
Pass-by-value eliminates a pointer indirection in the callee, which means it eliminates a load from memory. Godbolt:
int byvalue(std::string_view sv) { return sv.size(); }
int byref(const std::string_view& sv) { return sv.size(); }
---
byvalue:
movq %rsi, %rax
retq
byref:
movl 8(%rdi), %eax
retq
In the byvalue case, the string_view is passed in the register pair (%rdi, %rsi), so returning its “size” member is just a register-to-register move. In contrast, byref receives a reference to a string_view, passed in register %rdi, and has to do a memory load in order to extract the “size” member.
2. Eliminate a spill in the caller
When you pass by reference, the caller needs to put the thing’s address into a register. So the thing must have an address. Even if everything else in the caller could have been done with the thing in registers, the very act of passing the thing forces the caller to spill it onto the stack.
Pass-by-value eliminates the need to spill the argument, which sometimes means it eliminates the need for a stack frame in the caller at all. Godbolt:
int byvalue(std::string_view sv);
int byref(const std::string_view& sv);
void callbyvalue(std::string_view sv) { byvalue("hello"); }
void callbyref(std::string_view sv) { byref("hello"); }
---
.Lhello:
.asciz "hello"
callbyvalue:
movl $.Lhello, %edi
movl $5, %esi
jmp byvalue # tail call
callbyref:
subq $24, %rsp
movq $.Lhello, 8(%rsp)
movq $5, 16(%rsp)
leaq 8(%rsp), %rdi
callq byref
addq $24, %rsp
retq
In callbyvalue, we just set up the string_view argument’s data pointer and size member in %rdi and %rsi respectively, and then jump to byvalue. In callbyref, on the other hand, we need to pass the address of a string_view argument; so we make space on the stack. And then when byref returns, we need to clean up that space we made.
Previously on this blog: “It’s not always obvious when tail-call optimization is allowed” (2021-01-09).
3. Eliminate aliasing
When we pass by reference, we’re passing the callee a reference to an object they know nothing about. The callee doesn’t know who else might be holding a pointer to that object. The callee doesn’t know whether any of its own pointers might point at that object (or at any of its pieces). So the compiler must optimize the callee very conservatively, respecting those unknowns.
Pass-by-value gives the callee a brand-new copy of the object — a copy that definitely does not alias with any other object in the program. So the callee has greater opportunities for optimization. Godbolt:
void byvalue(std::string_view sv, size_t *p) {
*p = 0;
for (size_t i=0; i < sv.size(); ++i) *p += 1;
}
void byref(const std::string_view& sv, size_t *p) {
*p = 0;
for (size_t i=0; i < sv.size(); ++i) *p += 1;
}
---
byvalue:
movq %rsi, (%rdx)
retq
byref:
movq $0, (%rsi)
cmpq $0, 8(%rdi)
je .Lbottom
movl $1, %eax
.Ltop:
movq %rax, (%rsi)
leaq 1(%rax), %rcx
cmpq 8(%rdi), %rax
movq %rcx, %rax
jb .Ltop
.Lbottom:
retq
In byvalue, Clang is smart enough to see that a loop incrementing *p by 1, sv.size() times, starting from zero, is tantamount to a simple assignment *p = sv.size(). But in byref, Clang can’t make that same leap. Why not? Well, because byref needs to behave “correctly” even when called like this:
std::string_view sv = "hello";
size_t *size_p = &sv.__size_; // address of sv's "size" member
byref(sv, size_p);
In that situation, every increment of *size_p changes the result of sv.size(), causing the loop to run forever (or rather, until the value of sv.__size_ wraps around to zero and halts the loop). So the loop in byref, unlike the loop in byvalue, is not equivalent to a simple assignment! The compiler must generate machine code corresponding to its more complicated behavior.
byvalue doesn’t have to worry about that evil caller, because there is no (well-defined) way for the caller to pass in a copy of a string_view alongside a pointer that points into the interior of that copy.
In this example, we’re talking specifically about the possibility of aliasing targeting the string_view object’s own data members, not the characters viewed by it. Those characters, of course, can be aliased by pointers elsewhere in the program; but we’re not thinking about that in this specific example. Don’t let that confuse you!
To sum up:
C++ passes everything by value by default.
The pass-by-value default is optimal for small and cheap-to-copy types like int, char*, and pair.
Pass-by-value has at least three performance benefits, detailed above. But if the performance cost of making a copy outweighs all of these benefits — for large and/or expensive-to-copy types like string and vector — then prefer to pass by const reference. Pass-by-const-reference is an optimization of pass-by-value.
Small, trivially copyable, “parameter-only” types like C++17 string_view, C++20 span, and C++2b function_ref are explicitly designed to occupy the same category as int and char*. Pass them by value!
阅读(310) | 评论(0) | 转发(0) |
