[cfe-dev] Accessibility of name mangling utilities

[Anton Lokhmotov]

> -----Original Message-----
> From: Douglas Gregor [mailto:dgregor at apple.com]
> Sent: 12 May 2010 16:39
> To: Anton Lokhmotov
> Cc: cfe-dev at cs.uiuc.edu
> Subject: Re: [cfe-dev] Accessibility of name mangling utilities

> It would be a shame if you were forced to rewrite all of name mangling
> because of some limitations in the current implementation. 
The real limitation was exactly that the functionality was hidden inside
CodeGen.

> Could you explain why you had to re-implement name mangling for your task?
I was working on a prototype and just wanted to have something working in
under a day (hence, I wasn't forced to rewrite all of name mangling).  I
hope that when it comes to proper software engineering, the functionality
will be accessible.

BTW, perhaps it's time to bring up a limitation of the name mangling scheme
I found.  Namely, name mangling for vector functions doesn't take into
account the vector length.  For example, given declarations:

__attribute__((overloadable))  float8 select(float8,  float8,  int8);
__attribute__((overloadable)) float16 select(float16, float16, int16);

Clang produces only one LLVM declaration depending on whether in source code
a call to the 8-element version comes first or a call to the 16-element
version comes first.  If a call to the 8-element version comes first, the
declaration is:

declare <8 x float>
	@_Z6selectU8__vectorfS_U8__vectori(<8 x float>, <8 x float>, <8 x
i32>)

If, however, a call to the 16-element version comes first, the declaration
is:

declare <16 x float> 
	@_Z6selectU8__vectorfS_U8__vectori(<16 x float>, <16 x float>, <16 x
i32>)


In the former case, the calls look like:

; call for 8-element operands
%call8 = call <8 x float> @_Z6selectU8__vectorfS_U8__vectori(<8 x float>
%Vf8_0, <8 x float> %Vf8_1, <8 x i32> %Vi8) nounwind
...
; call for 16-element operands
%call16 = call <16 x float> bitcast (<8 x float> (<8 x float>, <8 x float>,
<8 x i32>)* @_Z6selectU8__vectorfS_U8__vectori to <16 x float> (<16 x
float>, <16 x float>, <16 x i32>)*)(<16 x float> %Vf16_0, <16 x float>
%Vf16_1, <16 x i32> %Vi16) nounwind


In the latter case, the calls look like:

; call for 16-element operands
%call16 = call <16 x float> @_Z6selectU8__vectorfS_U8__vectori(<16 x float>
%Vf16_1, <16 x float> %Vf16_2, <16 x i32> %Vi16)
...
; call for 8-element operands
%call8 = call <8 x float> bitcast (<16 x float> (<16 x float>, <16 x float>,
<16 x i32>)* @_Z6selectU8__vectorfS_U8__vectori to <8 x float> (<8 x float>,
<8 x float>, <8 x i32>)*)(<8 x float> %Vf8_0, <8 x float> %Vf8_1, <8 x i32>
%Vi8) nounwind


In either case, a call to the version that has no declaration involves
bitcast of the function pointer with unclear semantics.  Is this really as
bad as it seems?

Anton L.