[llvm-commits] [llvm] r114148 - in /llvm/trunk: lib/Analysis/ConstantFolding.cpp test/Transforms/InstCombine/fold-calls.ll

[Daniel Dunbar]

On Mon, Sep 20, 2010 at 1:27 PM, Dan Gohman <gohman at apple.com> wrote:
>
> On Sep 20, 2010, at 9:38 AM, Daniel Dunbar wrote:
>
>> Hi Dan,
>>
>> This change doesn't seem right to me. It is introducing a host ==
>> target dependency of sorts. Somehow this information should be coming
>> from the Target definitions or information in the IR file, not from
>> the host's fenv() implementation.
>
> There is a host == target dependence, but it's been there for many
> years, at a quick thumb through SVN history. My recent change just
> fixed a specific class of bugs. Other classes of bugs remain.

I guess adding a new configure check, plus System abstraction, plus
actual change, seems like a lot of work to put towards an inherently
broken model. Do we at least have a bug somewhere tracking a proper
fix for this?

 - Daniel

> The real way to fix this is to do something like implement sin, cos,
> etc. within APFloat.  GCC uses libmpfr to provide this functionality,
> for reference.
>
> Dan
>
>