4 ms·

Speaking as an Asm programmer for several decades: Calling conventions are stupid. They are the results of mindless stupid-compiler-oriented thinking from a tim

Speaking as an Asm programmer for several decades: Calling conventions are stupid. They are the results of mindless stupid-compiler-oriented thinking from a time when compilers produced horrible copy-paste-replace code. The CPU itself couldn't care less which registers you use for what. So many wasted bytes on moving values between registers, just because the calling convention wanted it there, and no other reason. The only need to pay attention to calling conventions is when you're interfacing with compiler-generated code. Modern CPUs are fast, but there's still tons of inefficiency in compiler output.

lmz 2y ago
> The only need to pay attention to calling conventions is when you're interfacing with compiler-generated code. So, the vast majority of code out there in the wild?
- userbinator 2y ago
  I mean it only matters at the interface.
- almostgotcaught 2y ago
  vast majority doesn't even begin to describe it - i would wager 10 years of my salary that the fraction of all currently running CPU instructions that were handwritten is so small that it's within the margin of error (i.e., random bit flips) for whatever computer you use to perform the count.
  - RandomBK 2y ago
    Depending on how you count, the ratio might not be that small. A lot of hot code are written in hand-coded inline assembly, so in terms of CPU cycles run it's probably non-negligible. i.e. take a look at the glibc implementation of 'strcmp` [0] [0] https://github.com/bminor/glibc/blob/master/sysdeps/x86_64/multiarch/strcmp-avx2.S#L204 https://github.com/bminor/glibc/blob/master/sysdeps/x86_64/m...
    - almostgotcaught 2y ago
      > A lot of hot code are written in hand-coded inline assembly I know... I write GPU assembly for a living... And still I make that wager. It's not a lot. It's not even a little. It's an epsilon (overall). And it gets smaller over time.
      - Kamq 2y ago
        What do you think the percentage of computation is with respect to hand-written vs compiler generated? If small hot loops tend to be disproportionately hand written, and certain programs spend the majority of time in their hot loops, that could still be a decent percentage of time/instructions executed.
    - lionkor 2y ago
      Now how much of that doesn't interface with compiler generated code?
- Spivak 2y ago
  If you're not interfacing with it, say linking it as a library, then it doesn't matter what you do.
  - lmz 2y ago
    Sure, but this sort of limits the kinds of thing you can realistically build unless you want to build everything from the ground up. Even in the case of code reuse with statically linked assembly files there would be some sort of "convention" about how to call and be called.
almostgotcaught 2y ago
do people think this is insightful? do you? > conventions are stupid all conventions are stupid when examined through the lens of an isolated island dweller. you might as well be saying something like "you only need to drive on the left-hand side of the road when you're driving on public roads".
- userbinator 2y ago
  Compilers were stupid, and that's how we ended up with this constant overhead of inefficiency long after they could've done better; it's only within the last decade or so that "custom calling conventions" started being even considered.
  - almostgotcaught 2y ago
    Calling conventions have no more to do with dumb or smart compilers than driving on the left-hand side of the road has to do with dumb or smart urban planners.
    - userbinator 2y ago
      Of course they do. An Asm programmer will naturally use the appropriate registers to minimise data movement (see also: PC BIOS interface - no stupid stack shit) depending on the circumstances, a stupid compiler will just push everything on the stack. A more intelligent compiler will behave more like the human programmer and decide how to pass parameters and save or restore registers on a case-by-case basis.
      - almostgotcaught 2y ago
        [flagged]
        josephg 2y ago
        I think you two are arguing past each other. Calling conventions are obviously needed for syscalls and dynamically linked libraries. I don't think anyone is denying that. But most function calls aren't made to shared code. Almost all function calls are made to private functions, which exist within a binary, and are under the control of the compiler. If I steelman the person you're arguing with, I think what they're claiming that adhering to any specific calling convention for internal functions results in a lot of dumb assembly. For example, imagine I have a C program where function a() calls b(). So long as b is confined to my binary, the "calling convention" of b doesn't matter. All that matters is that the compiler knows how to call it, and pass all the arguments. The compiler could jump to the function or call it. It could put parameters in registers or leave them on the stack. An awful lot of executed instructions exist to move parameters into the correct registers, save whatever was using those registers before the call, and restore them afterwards. If we imagine this simple C code: int do_stuff(int a, int b, int c, int d) { int i = 1000; i += func1(a, b); // a, b passed in rdi, rsi. i += func2(c, d); // Call to func2 overwrites rdi & rsi. return i; } If both func1 and func2 are forced to use the same calling convention, between calls to func1 and func2, the CPU needs to place c and d into whatever registers a and b were in a moment ago. But modern CPUs have lots of registers. If the compiler were more clever, it could just use different registers for the arguments of both functions and avoid shuffling everything around. Godbolt: https://c.godbolt.org/z/zqxbb447e https://c.godbolt.org/z/zqxbb447e (Aside: Its weird how different the assembly is between GCC and clang in this example!) But - moving values between registers (or between a register and the stack) is crazy fast anyway. I'd love to see some benchmarks showing how much of a difference this optimisation would make in practice.
timewizard 2y ago
> The CPU itself couldn't care less which registers you use for what. Not all registers encode as operands equivalently (implicit rdx:rax, implicit [rbx+al], limited [rbp/r13+imm8]). Some have other encoding restrictions or special purposes (rdi, rsi, rcx). When segmentation was a thing there were different default segments for each. Some are destroyed when certain opcodes used (syscall: rcx, r11). > So many wasted bytes on moving values between registers [...] Modern CPUs are fast Well, they've special cased this anyways, as these will often be caught in the rename stage and not even occupy an execution slot. Since we've long recognized that passing these values in registers instead of the stack is far more efficient, which is why the `fastcall` convention came about and got it's name way back in the x86 days. > but there's still tons of inefficiency in compiler output. Which is also why the 'inline' heuristic exists. In which case all of the calling conventions are fully abandoned. I mean, things like ELF dynamic symbol tables, and linux thread local storage annoy me far more than calling conventions ever have.
- userbinator 2y ago
  Well, they've special cased this anyways, as these will often be caught in the rename stage and not even occupy an execution slot They still need to be fetched and decoded, and take up space in caches and RAM that could be used for more purposeful instructions. Which is also why the 'inline' heuristic exists. Inlining has its own problems too. I mean, things like ELF dynamic symbol tables, and linux thread local storage annoy me far more than calling conventions ever have. Don't get me started on the whole ELF and dynamic linking situation...
antics 2y ago
Since no one seems to be pushing back I'll add my 2¢ here as a former compilers engineer. Calling conventions are just like any other style guide. Yes, any particular coding style is stupid, but it's still useful to have one that you are more or less committed to, especially if everyone else in the ecosystem is committed to it too. Frame pointers are a great example. Having a well-known and generic representation of %rbp is helpful when you go to use or integrate with existing tools like debuggers, link editors, or (say) most of the existing LLVM/GCC/whatever toolchain. Or when you want to expose a stable ABI to consumers for whatever reason (as, e.g., the Linux kernel famously does). Or, or, or. I think it's reasonable to say this has been mostly uncontroversial since at least the 90s. The discussion has changed a bit since (apparently) Go needed none of these things to succeed—not the LLVM compiler toolchain infrastructure, and also not the user-facing things like the debuggers. To hear Russ Cox tell the tale, this is mostly because they required flexibility, and I suppose they were right, since they did rewrite their linker 3 times, and sure enough, 15 years later, in the year of our lord 2024, most debugging in Go seems to happen by writing ASCII-shaped bytes to some kind of a file, somewhere, and then using the world's most expensive full text search engine to get those bytes so you can physically read them on a screen. A debugger does seem limited in use for that specific workflow, so maybe that was the right call, who knows. Anyway, I don't think there's ever been any real doubt that something like LLVM imposes a serious integration cost, but now that we have Go, the discussion has mostly shifted to "is it worth it", and seemingly the answer is "mostly yes" since nearly everyone building a new and hip native language uses LLVM or something like it. Every language is different, YMMV, etc., but I personally don't hear a lot of complaining about what a bummer it is that all these tools work pleasantly together instead of secretly sabotaging each other by loading up FP with whatever cursed data Go wanted to use it for. And why would they? What is more mysterious to me is how an actual assembly programmer came to defend Go's stance on ... assembly. Perhaps I'm the only one who reads these things, but at various points in the ASM docs[1] (which I will heretofore call "Mr Pike's wild ride") the author expresses a view that I think is reasonably well-described as "a tempered but pretty much open contempt for the practice as a whole". cf., > Instructions, registers, and assembler directives are always in UPPER CASE to remind you that assembly programming is a fraught endeavor. (Exception: the g register renaming on ARM.) Even if your feelings are hard to hurt though, if you ever crack open the toolchain and attempt to read the golang kind-of-IL-kind-of-x86, it is hard to walk away thinking "these people really get me and my profession". DI and FP are both normal registers! It uses the unicode interpunct instead of the plain old dot operator! It uses NIL instead of NULL! It is one thing to say calling conventions are stupid, but it's another thing entirely to give a great big hug to an almost-but-not-quite-assembly-code that is convenient neither for humans to type nor for tools to consume. [1]: https://go.dev/doc/asm https://go.dev/doc/asm
x-shadowban 2y ago
for functions that don't escape the current compilation unit (`static` functions, anonymous namespace functions), can/do compilers ignore calling conventions and do the faster thing? Of course, they can just inline, and that makes this moot.

lmz 2y ago

> The only need to pay attention to calling conventions is when you're interfacing with compiler-generated code. So, the vast majority of code out there in the wild?

userbinator 2y ago
I mean it only matters at the interface.
almostgotcaught 2y ago
vast majority doesn't even begin to describe it - i would wager 10 years of my salary that the fraction of all currently running CPU instructions that were handwritten is so small that it's within the margin of error (i.e., random bit flips) for whatever computer you use to perform the count.
- RandomBK 2y ago
  Depending on how you count, the ratio might not be that small. A lot of hot code are written in hand-coded inline assembly, so in terms of CPU cycles run it's probably non-negligible. i.e. take a look at the glibc implementation of 'strcmp` [0] [0] https://github.com/bminor/glibc/blob/master/sysdeps/x86_64/multiarch/strcmp-avx2.S#L204 https://github.com/bminor/glibc/blob/master/sysdeps/x86_64/m...
  - almostgotcaught 2y ago
    > A lot of hot code are written in hand-coded inline assembly I know... I write GPU assembly for a living... And still I make that wager. It's not a lot. It's not even a little. It's an epsilon (overall). And it gets smaller over time.
    - Kamq 2y ago
      What do you think the percentage of computation is with respect to hand-written vs compiler generated? If small hot loops tend to be disproportionately hand written, and certain programs spend the majority of time in their hot loops, that could still be a decent percentage of time/instructions executed.
  - lionkor 2y ago
    Now how much of that doesn't interface with compiler generated code?
Spivak 2y ago
If you're not interfacing with it, say linking it as a library, then it doesn't matter what you do.
- lmz 2y ago
  Sure, but this sort of limits the kinds of thing you can realistically build unless you want to build everything from the ground up. Even in the case of code reuse with statically linked assembly files there would be some sort of "convention" about how to call and be called.

almostgotcaught 2y ago

do people think this is insightful? do you? > conventions are stupid all conventions are stupid when examined through the lens of an isolated island dweller. you might as well be saying something like "you only need to drive on the left-hand side of the road when you're driving on public roads".

userbinator 2y ago
Compilers were stupid, and that's how we ended up with this constant overhead of inefficiency long after they could've done better; it's only within the last decade or so that "custom calling conventions" started being even considered.
- almostgotcaught 2y ago
  Calling conventions have no more to do with dumb or smart compilers than driving on the left-hand side of the road has to do with dumb or smart urban planners.
  - userbinator 2y ago
    Of course they do. An Asm programmer will naturally use the appropriate registers to minimise data movement (see also: PC BIOS interface - no stupid stack shit) depending on the circumstances, a stupid compiler will just push everything on the stack. A more intelligent compiler will behave more like the human programmer and decide how to pass parameters and save or restore registers on a case-by-case basis.
    - almostgotcaught 2y ago
      [flagged]
      - josephg 2y ago
        I think you two are arguing past each other. Calling conventions are obviously needed for syscalls and dynamically linked libraries. I don't think anyone is denying that. But most function calls aren't made to shared code. Almost all function calls are made to private functions, which exist within a binary, and are under the control of the compiler. If I steelman the person you're arguing with, I think what they're claiming that adhering to any specific calling convention for internal functions results in a lot of dumb assembly. For example, imagine I have a C program where function a() calls b(). So long as b is confined to my binary, the "calling convention" of b doesn't matter. All that matters is that the compiler knows how to call it, and pass all the arguments. The compiler could jump to the function or call it. It could put parameters in registers or leave them on the stack. An awful lot of executed instructions exist to move parameters into the correct registers, save whatever was using those registers before the call, and restore them afterwards. If we imagine this simple C code: int do_stuff(int a, int b, int c, int d) { int i = 1000; i += func1(a, b); // a, b passed in rdi, rsi. i += func2(c, d); // Call to func2 overwrites rdi & rsi. return i; } If both func1 and func2 are forced to use the same calling convention, between calls to func1 and func2, the CPU needs to place c and d into whatever registers a and b were in a moment ago. But modern CPUs have lots of registers. If the compiler were more clever, it could just use different registers for the arguments of both functions and avoid shuffling everything around. Godbolt: https://c.godbolt.org/z/zqxbb447e https://c.godbolt.org/z/zqxbb447e (Aside: Its weird how different the assembly is between GCC and clang in this example!) But - moving values between registers (or between a register and the stack) is crazy fast anyway. I'd love to see some benchmarks showing how much of a difference this optimisation would make in practice.

timewizard 2y ago

> The CPU itself couldn't care less which registers you use for what. Not all registers encode as operands equivalently (implicit rdx:rax, implicit [rbx+al], limited [rbp/r13+imm8]). Some have other encoding restrictions or special purposes (rdi, rsi, rcx). When segmentation was a thing there were different default segments for each. Some are destroyed when certain opcodes used (syscall: rcx, r11). > So many wasted bytes on moving values between registers [...] Modern CPUs are fast Well, they've special cased this anyways, as these will often be caught in the rename stage and not even occupy an execution slot. Since we've long recognized that passing these values in registers instead of the stack is far more efficient, which is why the `fastcall` convention came about and got it's name way back in the x86 days. > but there's still tons of inefficiency in compiler output. Which is also why the 'inline' heuristic exists. In which case all of the calling conventions are fully abandoned. I mean, things like ELF dynamic symbol tables, and linux thread local storage annoy me far more than calling conventions ever have.

userbinator 2y ago
Well, they've special cased this anyways, as these will often be caught in the rename stage and not even occupy an execution slot They still need to be fetched and decoded, and take up space in caches and RAM that could be used for more purposeful instructions. Which is also why the 'inline' heuristic exists. Inlining has its own problems too. I mean, things like ELF dynamic symbol tables, and linux thread local storage annoy me far more than calling conventions ever have. Don't get me started on the whole ELF and dynamic linking situation...

antics 2y ago

Since no one seems to be pushing back I'll add my 2¢ here as a former compilers engineer. Calling conventions are just like any other style guide. Yes, any particular coding style is stupid, but it's still useful to have one that you are more or less committed to, especially if everyone else in the ecosystem is committed to it too. Frame pointers are a great example. Having a well-known and generic representation of %rbp is helpful when you go to use or integrate with existing tools like debuggers, link editors, or (say) most of the existing LLVM/GCC/whatever toolchain. Or when you want to expose a stable ABI to consumers for whatever reason (as, e.g., the Linux kernel famously does). Or, or, or. I think it's reasonable to say this has been mostly uncontroversial since at least the 90s. The discussion has changed a bit since (apparently) Go needed none of these things to succeed—not the LLVM compiler toolchain infrastructure, and also not the user-facing things like the debuggers. To hear Russ Cox tell the tale, this is mostly because they required flexibility, and I suppose they were right, since they did rewrite their linker 3 times, and sure enough, 15 years later, in the year of our lord 2024, most debugging in Go seems to happen by writing ASCII-shaped bytes to some kind of a file, somewhere, and then using the world's most expensive full text search engine to get those bytes so you can physically read them on a screen. A debugger does seem limited in use for that specific workflow, so maybe that was the right call, who knows. Anyway, I don't think there's ever been any real doubt that something like LLVM imposes a serious integration cost, but now that we have Go, the discussion has mostly shifted to "is it worth it", and seemingly the answer is "mostly yes" since nearly everyone building a new and hip native language uses LLVM or something like it. Every language is different, YMMV, etc., but I personally don't hear a lot of complaining about what a bummer it is that all these tools work pleasantly together instead of secretly sabotaging each other by loading up FP with whatever cursed data Go wanted to use it for. And why would they? What is more mysterious to me is how an actual assembly programmer came to defend Go's stance on ... assembly. Perhaps I'm the only one who reads these things, but at various points in the ASM docs[1] (which I will heretofore call "Mr Pike's wild ride") the author expresses a view that I think is reasonably well-described as "a tempered but pretty much open contempt for the practice as a whole". cf., > Instructions, registers, and assembler directives are always in UPPER CASE to remind you that assembly programming is a fraught endeavor. (Exception: the g register renaming on ARM.) Even if your feelings are hard to hurt though, if you ever crack open the toolchain and attempt to read the golang kind-of-IL-kind-of-x86, it is hard to walk away thinking "these people really get me and my profession". DI and FP are both normal registers! It uses the unicode interpunct instead of the plain old dot operator! It uses NIL instead of NULL! It is one thing to say calling conventions are stupid, but it's another thing entirely to give a great big hug to an almost-but-not-quite-assembly-code that is convenient neither for humans to type nor for tools to consume. [1]: https://go.dev/doc/asm https://go.dev/doc/asm

x-shadowban 2y ago

for functions that don't escape the current compilation unit (`static` functions, anonymous namespace functions), can/do compilers ignore calling conventions and do the faster thing? Of course, they can just inline, and that makes this moot.