# A footnote on “Three reasons to pass std::string_view by value”

Several readers have responded to my recent post “Three reasons to pass std::string_view by value” (2021-11-09), pointing out that (while everything I said is true of the Itanium x86-64 ABI used on Linux and Mac) unfortunately Microsoft’s x86-64 ABI does not pass std::string_view in registers, but instead passes it by “hidden pointer,” so that at the machine level, you get basically the same codegen as if you had been passing by reference all along. Except that actually you’re passing a copy by reference, so you get one extra 16-byte copy in addition to all the other stack traffic MSVC is doing.

This is a big problem baked into Microsoft’s x86-64 ABI, and nerfs a lot of the performance benefits mentioned in my post. However, my advice remains the same: Even if your primary platform today is Windows, you should habitually pass std::string_view by value!

Let’s see how, exactly, MSVC demolishes the performance gains listed in my previous post.

## 1. Eliminate a pointer indirection in the callee?

Microsoft’s ABI means that we don’t eliminate indirection in the callee when we pass class types (of size >8) by value. Godbolt:

int byvalue(std::string_view sv) { return sv.size(); }

int byref(const std::string_view& sv) { return sv.size(); }

---

byvalue:
mov eax, DWORD PTR [rcx+8]
ret 0

byref:
mov eax, DWORD PTR [rcx+8]
ret 0


Because Microsoft’s x86-64 ABI passes by hidden pointer, we end up with the same x86-64 code in both cases.

However, just to illustrate that this is a performance issue with the Microsoft x86-64 ABI, and not with “Visual Studio” in general, in the Godbolt above I also included MSVC’s output for the 64-bit ARM processor. On ARM, Microsoft follows a smarter calling convention modeled after the Itanium ABI, and does pass string_view in registers. So in the unlikely case that you dual-compile for Windows on ARM, you should definitely pass string_view by value!

## 2. Eliminate a spill in the caller?

Microsoft’s ABI means that we don’t eliminate the spill. 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:
sub rsp, 56
mov QWORD PTR $T2[rsp+8], 5 lea rax, OFFSET FLAT:.Lhello mov QWORD PTR$T2[rsp], rax
lea rcx, QWORD PTR $T1[rsp] movaps xmm0, XMMWORD PTR$T2[rsp]
movdqa XMMWORD PTR $T1[rsp], xmm0 call byvalue add rsp, 56 ret 0 callbyref: sub rsp, 56 lea rax, OFFSET FLAT:.Lhello mov QWORD PTR$T1[rsp+8], 5
lea rcx, QWORD PTR $T1[rsp] mov QWORD PTR$T1[rsp], rax
call byref
ret 0


You can see the extra 16-byte copy happening in callbyvalue: that’s the movaps/movdqa instruction pair. But you can also see that there’s no reason for MSVC to be doing that — its optimizer should have coalesced those temporaries! In general, this extra copy can cause MSVC’s codegen for pass-by-value to be worse than the codegen for pass-by-reference; but in this specific case, I’d just call this a missed optimization and file a bug against the compiler. There’s no reason either of these functions needs a 56-byte stack frame. (And notice: passing by reference doesn’t reduce the size of MSVC’s stack frame! This codegen is just abymally bad and I don’t know why.) This is with -O2, by the way.

Again, my Godbolt includes MSVC’s codegen for 64-bit ARM, showing that you do get this performance benefit when you pass by value on ARM. Even on ARM, MSVC’s codegen quality is worse than I’d expect; but the pass-by-value version does clearly win out.

## 3. Eliminate aliasing?

Having seen a sample of MSVC’s optimization smarts in the previous section, you should be unsurprised to learn that MSVC doesn’t seem to feed its optimizer with aliasing information. 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:
mov r8, QWORD PTR [rcx+8]
xor eax, eax
mov QWORD PTR [rdx], rax
test r8, r8
je .Lbottom
.Ltop:
inc rax
mov QWORD PTR [rdx], rax
sub r8, 1
jne .Ltop
.Lbottom:
ret 0

byref:
xor eax, eax
mov QWORD PTR [rdx], rax
cmp QWORD PTR [rcx+8], rax
jbe .Lbottom
mov r8d, eax
.Ltop:
inc r8
inc rax
mov QWORD PTR [rdx], r8
cmp rax, QWORD PTR [rcx+8]
jb .Ltop
.Lbottom:
ret 0


Because this is a Visual Studio optimizer issue and not a Windows x64 ABI issue, this one applies equally to MSVC-on-x64 and MSVC-on-ARM. The only way to claw back this performance, as far as I know, is to switch to Clang or GCC.

So, given that string_view’s pass-by-value codegen and its pass-by-reference codegen are pretty much equally awful on MSVC, why do I still specifically recommend pass-by-value?

• Even on Windows, where pass-by-value doesn’t gain you boatloads of performance, it still doesn’t lose any performance to speak of. (Yes, there’s that 16-byte copy we observed in section 2; but you’re hitting the stack either way, so it’s the difference between one L1 cache hit or two. I don’t think you’ll notice, what with all the other stuff going on in this post.)

• Passing string_view by value is unambiguously the right thing to do, if your code is ever going to run on x86-64 Linux or Mac. You don’t want to hard-code a bunch of pass-by-reference and then have to undo it later.

• Even if you’re a 100% Visual Studio shop, passing string_view by value is unambiguously the right thing to do, if your code is ever going to run on ARM or any other non-x86 architecture. Microsoft’s wonky ABI applies only to Windows on x86-64, not to Windows on other architectures.

• If you’re writing (possibly open-source) library code that might one day be copy-pasted onto a non-Windows-x86-64 platform, then again you should pass string_view by value, so you don’t have to undo it later.

• If you ever show your code to any other human, they’re likely to ask you, “Hey, why are you passing by reference here? Wouldn’t it be more idiomatic to pass by value?” In general, if there are two ways of doing something, and their performance is equivalent, but one way is idiomatic and the other way is idiosyncratic, you should prefer the idiomatic way.

To repeat the bottom line from my previous post:

• 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!
Posted 2021-11-19