Handmade Hero

Right, that is true. Although if you are using handmade-like build, where everything is single-translation unit, then /GL doesn't really matter. There are no optimizations across translation-units, because there is only one translation unit. So you won't gain anything from link-time optimizations. Everything can be optimized during compilation with /O2 or /Ox /Ot.

Greetings everyone, I know that it's quite an old thread but, I've noticed it just a few days ago and since then I'm trying as hard as I can to implement the casting functions by myself with inline assembly. I think I'm pretty much done with the _ftoui3 (float to unsigned int casting) function, even though I haven't done any exception checking and I don't know how efficient it is, it seems to be working quite nicely. You can judge for yourself:



However, I got stuck on the _dtoui3 (double to unsigned int casting) function. You see, the problem is that although there are indeed instructions that can truncate floating point numbers successfully, they all share a common unfortunate drawback: They all store the truncated value as a signed integer, which will lead to some data loss when a value stored in a double exceeds the limit of a signed integer. I've scanned the whole Internet trying to find different instructions or different methods to tackle this problem but to no avail.
If anyone of you could share your wisdom with me and shed some light on this minor topic, I would appreciate it very much. I'm certain that there is a way since the CRT managed to implement it somehow, but the disassembly of this function looks very intimidating so I can't figure out anything out of it.

I found this instruction, but it seems to be AVX512. https://www.felixcloutier.com/x86/VCVTTSD2USI.html

SedatedSnail

I found this instruction, but it seems to be AVX512. https://www.felixcloutier.com/x86/VCVTTSD2USI.html

Oh man, that's so close!!!
Unfortunately, there aren't many people yet (myself included) who got the necessary processor to support this instruction set :(

Using AVX512 instruction when you are compiling for 32-bit code seems very wrong... :)

Typically you implement _ftoui3 like this:



if statement does not need to be real "if", it can be conditional move.

This is how it would look like with SSE2:


Not sure if this will handle all the Inifinty or NaN floats exactly same as your regular C runtime cast, but otherwise it should work exactly the same.

Similar approach works four doubles in 32-bit code.

If you want to avoid SSE instructions, you can probably us x87 fpu instructions in similar way, or alternatively you can extract mantissa & exponents bits and use them to calculate actual value yourself. Check the code in llvm compiler-rt library that does this: https://github.com/llvm-mirror/co.../builtins/fp_fixuint_impl.inc#L17

Awesome!!!
Thank you very much my good man, I'll use this information wisely!

mmozeiko

Similar approach works four doubles in 32-bit code.

How exactly is it similar if the truncation instruction returns a signed value, and as a result, causes you to lose precision if your number exceeds the signed integer limit, which can happen if you work with doubles.
You need to find another instruction that truncates and returns an unsigned value or implement the truncation yourself somehow, don't you?

It will work exactly the same:


Or in asm:


(NOTE: I have not actually tested this code, so verify that it works yourself)

The difference is only in conversion operation - cvttss2si vs cvttsd2si. Both returns 32-bit signed int, which you fixup, if it is "negative".

If you are running in 64-bit code, then you can use just one instruction: "cvttsd2si rax, xmm0".

Very late to the discussion, I have waited an episode to go through all this.

Thank you so much mmozeiko, fantastic work!!!
:) :)

Hi!

Firstly, there are some great posts here and I've learned a lot about dropping the CRT.

But I'm having a problem implementing a DLL with pure virtual base classes - I always get this error

error LNK2001: unresolved external symbol "const type_info::`vftable'" ([email protected]@[email protected])

I need to call into the class via the base object (from an exe) but I can't get past this problem :-(

Can anyone help resolve this?



Incidentally, I cant link to ANY .libs - I need to be able to compile a DLL on users machines that don't have the windows SDK installed (It's for a scripting system). Everything works apart from the base class thing!

Hey, it's been a long time since this thread was published.

I'm working around it right know and I got into some trouble here.

I'm using clang-cl as the compiler. Linker still is the link.exe from MSVC. Everything went fine until I got to the part where we need to define memset and memcpy and use #pragma function to tell the compiler not inline them. It seems that Clang does not knows what #pragma function is, it says "error: unknown pragma ignored".

Does anyone knows a work around for this? How to tell Clang to do the same as MSVC does with #pragma function? Have anyone tried to avoid the C Runtime using Clang?

Also... another thing that I noticed is that, the linker generates a .lib file and a .exp file, don't know why. Without /NODEFAULTLIB it doesn't generates this files. Don't know if it's because I'm using Clang, which don't makes sense since who generates these files is the linker, and the linker is link.exe from MSVC. Is there a switch to stop generating these files? Is it even possible?

Thanks for now!

EDIT: I solved this problem:

I found that I wrote the memset and memcpy function signatures with static (aka internal from Handmade Hero) by force of habit, since I also use a unity build system and follow what Casey believes, maybe telling the compiler that all functions are static, compilation is faster.

Don't even needed any #pragma, nothing.

Now, the .lib and .exp files are gone too! I don't have that much timing programming in C and extern and static are keywords that I don't master yet exactly what they do when it comes to linking.