[llvm-commits] [llvm-gcc-4.2] r82640 - /llvm-gcc-4.2/trunk/gcc/llvm-convert.cpp

[Dan Gohman]

Using an intrinsic to convey "this function need not set errno" seem
like a very heavy-weight solution. Wouldn't it be better to make use
of the "readonly" or "readnone" function attributes? A "sqrt" call
with "readonly" can be known to not set errno.

The intention of the "like the libm sqrt functions would" language was
to support vector forms of these intrinsics, which should behave the
same way as the scalar forms.

Dan

On Sep 23, 2009, at 2:18 PM, Dale Johannesen wrote:

> While this is certainly necessary given the current IR definition, I'm
> not too happy about it, as we will lose something because the
> optimizers must consider that sqrt might set errno.  What I'd like to
> do is redefine llvm.sqrt in the IR to mean "same behavior as sqrt
> except that it doesn't set errno", and likewise for llvm.pow, etc.  It
> looks like that was the original intent, actually, but that's not what
> it says.  Any objections?
>
> On Sep 23, 2009, at 11:50 AMPDT, Dale Johannesen wrote:
>
>> Author: johannes
>> Date: Wed Sep 23 13:50:32 2009
>> New Revision: 82640
>>
>> URL: http://llvm.org/viewvc/llvm-project?rev=82640&view=rev
>> Log:
>> The result of sqrt(negative) is implementation-defined, but follows
>> IEEE754 in most current implementations. llvm.sqrt, which has
>> undefined
>> behavior for such inputs, is an inappropriate substitute which can
>> change
>> the result.
>
> _______________________________________________
> llvm-commits mailing list
> llvm-commits at cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits