llvm dev - Jan 2012 - [LLVMdev] Implementation of builtins/intrinsics

Hi,

Lately, I have been wondering how a compiler like GCC or Clang implements
builtin functions like sqrt, sin, and also more bit-specific things like
bcmp and ffs. Originially, these were all deferred to the C library where
they might still have a "backup implementation", but I don't
understand why
a compiler would choose the non-builtin version, as it should know its
version is faster (or that's the way it should be at least).

 - How does LLVM cope with builtin's that are target specific? Does the
Target backend know of all these specialized instructions?
 - Does LLVM have a representation for all C library functions that have or
might have a future hardware implementation? Or is this mostly handled by
the whole toolchain (Clang adds target-specific info into the LLVM bitcode
which is forced to produce the asm instruction if the target supports it)?
 - Is it possible to code a sqrt or sin in LLVM bitcode that will use the
hardware instruction if it's compiled for a certain target?

I know these are probably questions born out of misunderstanding, but
that's why I'm asking. It's important to know what LLVM could
provide if I
say were to always bypass the system C library for these "builtins".

Thanks!

Ruben
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20120120/afb69aef/attachment.html>

Le 20 janv. 2012 à 14:38, Ruben Van Boxem a écrit :
> Hi,
> 
> Lately, I have been wondering how a compiler like GCC or Clang implements
builtin functions like sqrt, sin, and also more bit-specific things like bcmp
and ffs. Originially, these were all deferred to the C library where they might
still have a "backup implementation", but I don't understand why a
compiler would choose the non-builtin version, as it should know its version is
faster (or that's the way it should be at least).
> 
>  - How does LLVM cope with builtin's that are target specific? Does the
Target backend know of all these specialized instructions?
>  - Does LLVM have a representation for all C library functions that have or
might have a future hardware implementation? Or is this mostly handled by the
whole toolchain (Clang adds target-specific info into the LLVM bitcode which is
forced to produce the asm instruction if the target supports it)?
>  - Is it possible to code a sqrt or sin in LLVM bitcode that will use the
hardware instruction if it's compiled for a certain target?

See the LLVM language reference for the list of supported intrinsics:
http://llvm.org/docs/LangRef.html#int_libc

-- Jean-Daniel

2012/1/20 Jean-Daniel Dupas <devlists at shadowlab.org>
>
> Le 20 janv. 2012 à 14:38, Ruben Van Boxem a écrit :
>
> > Hi,
> >
> > Lately, I have been wondering how a compiler like GCC or Clang
> implements builtin functions like sqrt, sin, and also more bit-specific
> things like bcmp and ffs. Originially, these were all deferred to the C
> library where they might still have a "backup implementation",
but I don't
> understand why a compiler would choose the non-builtin version, as it
> should know its version is faster (or that's the way it should be at
least).
> >
> >  - How does LLVM cope with builtin's that are target specific?
Does the
> Target backend know of all these specialized instructions?
> >  - Does LLVM have a representation for all C library functions that
have
> or might have a future hardware implementation? Or is this mostly handled
> by the whole toolchain (Clang adds target-specific info into the LLVM
> bitcode which is forced to produce the asm instruction if the target
> supports it)?
> >  - Is it possible to code a sqrt or sin in LLVM bitcode that will use
> the hardware instruction if it's compiled for a certain target?
>
>
> See the LLVM language reference for the list of supported intrinsics:
> http://llvm.org/docs/LangRef.html#int_libc

Thanks for that. I still wonder what happens to stuff like log1p, which
translates to a machine instruction on x86. Can the LLVM backend discover
that this operation is called for, and use the native instruction in this
case?

Thanks,

Ruben

>
> -- Jean-Daniel
>
>
>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20120121/0e63bd7e/attachment.html>