[llvm-commits] [llvm-gcc-4.2] r55796 - /llvm-gcc-4.2/trunk/gcc/llvm-convert.cpp

[Duncan Sands]

Hi Dale,

> OK, I do not know Ada.   It seems that its...interesting...FP model  
> would need more than I have suggested, at least to do a good job.

it has this model because it targets (or used to target) machines
which do not support IEEE floating point.  That's unfortunate because
the IEEE is clearly superior.

> As for errno, the fact that Linux uses it is a target property and  
> again it does not belong in the .bc file.  That is, we do not need  
> errno and non-errno versions of the intrinsics.  (Do you know any real  
> code that uses errno btw?  I've never seen it outside test suites, and  
> it's generally recognized as a poor design.)

OK, I thought you wanted to only do simplifications on intrinsics,
which would require an intrinsic for "pow with errno".  I agree that
errno sucks.  Ada may have a crazy fp model but at least it doesn't
have errno :)

> > Yes, it would be useful, but as a first step I would rather have
> > llvm(-gcc) produce correct code, and worry about optimization later.
> 
> I know, and I'm going along with it, with serious misgivings.  In what  
> way do you think we don't have this?

I already sent a testcase which shows that codegen does CSE on
"readonly" math functions as if they were readnone.  That's the
only incorrectness that I know of.

Ciao,

Duncan.