[llvm-commits] [llvm] r73243 - in /llvm/trunk: lib/Transforms/Scalar/InstructionCombining.cpp test/Transforms/InstCombine/fsub-fsub.ll

[Chris Lattner]

On Jun 12, 2009, at 2:07 PM, Dale Johannesen wrote:

>
> On Jun 12, 2009, at 1:52 PMPDT, Dan Gohman wrote:
>> InstCombine doesn't currently know about UnsafeFPMath.
>>
>> DAGCombine does know about it, and it already does this
>> transformation.
>
> OK.  Why on earth doesn't InstCombine know about it?  That seems
> really wrong, there's a lot of target-independent optimizations that
> should be controlled by this.

The optimizer (unlike codegen) doesn't have "global variables" that  
affect interpretation of the IR.  To do this xform in instcombine we  
need to enhance the various FP instructions to have flags that  
indicate (on a per operation basis) whether it is ok to assume  
"unsafeness", etc.  Same thing for rounding modes etc.

-Chris