Intrinsics without builtin

Clang Frontend
#1 
Hi,

I got a solution to use intrinsics without the need of builtin functions from:
[http://wiki.llvm.org/HowTo:_Call_LLVM_intrinsics_from_C_code](http://wiki.llvm.org/HowTo:_Call_LLVM_intrinsics_from_C_code)

Do you know if this approach is still valid ?

The example is the following:

  extern "C" int llvm_cas(volatile int*, int, int) asm("llvm.atomic.cmp.swap.i32");

 
  int foo(volatile int* add, int from, int to) {                                  

      return llvm_cas(add, from, to);                                             
  }

And the expected llvm output is:

  define i32 @_Z3fooPViii(i32* %add, i32 %from, i32 %to) {
  entry:
  	%0 = tail call i32 @llvm.atomic.cmp.swap.i32(i32* %add, i32 %from, i32 %to); <i32> [#uses=1]
  	ret i32 %0
  }

  declare i32 @llvm.atomic.cmp.swap.i32(i32*, i32, i32)

However, when I compile the same code, it generates the following call:

   %call = tail call i32 @"\01llvm.atomic.cmp.swap.i32"(i32* %add, i32 %from, i32 %to) nounwind optsize

which looks to me a regular call and not a direct call to the intrinsic.

   Thank you,

  Damien
#2

Hi,

I got a solution to use intrinsics without the need of builtin functions from:
http://wiki.llvm.org/HowTo:_Call_LLVM_intrinsics_from_C_code

...

However, when I compile the same code, it generates the following call:

   %call = tail call i32 @"\01llvm.atomic.cmp.swap.i32"(i32* %add, i32
%from, i32 %to) nounwind optsize

which looks to me a regular call and not a direct call to the intrinsic.

Hi Damien,

I had the same problem a while back, please see if the attached patches work
for you (apply against clang trunk).

(In the end I decided on a different solution to my problem, which
is why I never sent these patches for approval.)

Thanks,

0001-Do-not-use-IR-marker-for-LLVM-intrinsics.patch (1.18 KB)

0002-If-this-is-an-intrinsic-function-set-the-function-s-.patch (1.17 KB)

#3

Thank you for the patch. Though I finally decided to go for builtins (this seems to be a cleaner solution).
My worry is that I don’t know how standard, compiler dependant is the following declaration:

extern int llvm_cas(volatile int*, int, int) asm("llvm.atomic.cmp.swap.i32");

Do you have any idea of the pros/cons of using either builtins or the previous declaration ?

Thank you Peter !

Damien

#4

Thank you for the patch. Though I finally decided to go for builtins (this seems to be a cleaner solution).
My worry is that I don't know how standard, compiler dependant is the following declaration:
extern int llvm_cas(volatile int*, int, int) asm("llvm.atomic.cmp.swap.i32");

That's about as compiler dependent as you can get.

Do you have any idea of the pros/cons of using either builtins or the previous declaration ?

Use documented APIs if at all possible.

-eric