3 ms·

I would argue that inline assembly, while not in the standard, is really just a convenience feature -- it is part of the standard to declare an extern reference

I would argue that inline assembly, while not in the standard, is really just a convenience feature -- it is part of the standard to declare an extern reference to a function in the symbol table and jump to it, it just requires a separate ASM object to link alongside your C object. Inline assembly doesn't allow anything you can't do without it.

convolvatron 2mo ago
that's kind of not true. inline asm lets me refer to the register that the compiler placed a value in. lets say I really want to use popcnt in my inner loop. with inline assembly I can just shove it in there. external linkage forces a function call overhead that can't be inlined, which obviates any benefit I might have had from using the specialized instruction.
- treyd 2mo ago
  You'd probably also be able to use an intrinsic to avoid the pain of inline asm and make it portable.
  - jonhohle 2mo ago
    Intrinsics usually come after new instructions have existed long enough for the higher level pattern across several architectures to establish a common pattern. If specific hardware is being targeted you may need to use those instructions before intrinsics exist.
  - convolvatron 2mo ago
    intrinsics are nicer in every way, assuming they exist. but some instructions are inherently non-portable. performance instruction like my popcnt example are good candidates since they can be implemented at varying costs on other architectures. but for systems programming there are control register and mode switch instructions that really aren't. some of those can be put into separate asm routines, but there are some that are poorly suited. segment long jumps on x86 are maybe an example. rdtsc is another one potentially. its also true that when I unwrap my new spin with fancy new instructions its unlikely to have a robust set of instrinsics around them. inline asm is a real mess, I always regret tussling with it, but its kind of pragmatically necessary if you're actually working at the metal in a high performance or embedded context unless you're doing the whole thing in assembly.
jonhohle 2mo ago
What does an external function have to do with the C supporting or not supporting ASM? The symbol is just a symbol from another object. That could be written in any language that follows the ABI (not that ASM has any enforcement of ABi to begin with). The symbol resolves to an address and nothing more. Saying inline ASM is no different than a function call is like saying standard control structures are no different from function calls. I suppose from a Smalltalk perspective that could be true, but is that the mental model most programmers use? I work on a system from the 90s with custom instructions. GNU-as was patched to understand the instructions. They’re used through macros that ultimately expand to inline ASM. Without this, you’d need function inlining, which may or may not be possible with a linked object (it certainly wasn’t standard in the 90s). So now a single instruction turns into stack management, a jump, more stack management and a return. At that point any benefit to a specialized instruction may be erased, or in the case I’m dealing with talking to external hardware becomes unreasonably expensive.
Dwedit 2mo ago
Inline assembly is inline. You're not following the platform ABI's calling convention here, you are choosing input registers, output registers, and trashed registers right there. Following the platform's ABI and carrying out the function call has a cost, merely picking registers does not.

convolvatron 2mo ago

that's kind of not true. inline asm lets me refer to the register that the compiler placed a value in. lets say I really want to use popcnt in my inner loop. with inline assembly I can just shove it in there. external linkage forces a function call overhead that can't be inlined, which obviates any benefit I might have had from using the specialized instruction.

treyd 2mo ago
You'd probably also be able to use an intrinsic to avoid the pain of inline asm and make it portable.
- jonhohle 2mo ago
  Intrinsics usually come after new instructions have existed long enough for the higher level pattern across several architectures to establish a common pattern. If specific hardware is being targeted you may need to use those instructions before intrinsics exist.
- convolvatron 2mo ago
  intrinsics are nicer in every way, assuming they exist. but some instructions are inherently non-portable. performance instruction like my popcnt example are good candidates since they can be implemented at varying costs on other architectures. but for systems programming there are control register and mode switch instructions that really aren't. some of those can be put into separate asm routines, but there are some that are poorly suited. segment long jumps on x86 are maybe an example. rdtsc is another one potentially. its also true that when I unwrap my new spin with fancy new instructions its unlikely to have a robust set of instrinsics around them. inline asm is a real mess, I always regret tussling with it, but its kind of pragmatically necessary if you're actually working at the metal in a high performance or embedded context unless you're doing the whole thing in assembly.

jonhohle 2mo ago

What does an external function have to do with the C supporting or not supporting ASM? The symbol is just a symbol from another object. That could be written in any language that follows the ABI (not that ASM has any enforcement of ABi to begin with). The symbol resolves to an address and nothing more. Saying inline ASM is no different than a function call is like saying standard control structures are no different from function calls. I suppose from a Smalltalk perspective that could be true, but is that the mental model most programmers use? I work on a system from the 90s with custom instructions. GNU-as was patched to understand the instructions. They’re used through macros that ultimately expand to inline ASM. Without this, you’d need function inlining, which may or may not be possible with a linked object (it certainly wasn’t standard in the 90s). So now a single instruction turns into stack management, a jump, more stack management and a return. At that point any benefit to a specialized instruction may be erased, or in the case I’m dealing with talking to external hardware becomes unreasonably expensive.

Dwedit 2mo ago

Inline assembly is inline. You're not following the platform ABI's calling convention here, you are choosing input registers, output registers, and trashed registers right there. Following the platform's ABI and carrying out the function call has a cost, merely picking registers does not.