[llvm-commits] Patch: Fast-math flags for intrinsics

[Chris Lattner]

On Dec 18, 2012, at 2:07 PM, Michael Ilseman <milseman at apple.com> wrote:

> Attached are patches for fast-math support for IntrinsicInst. This includes bitcode changes to separate out IntrinsicInsts from CallInsts.
> 
> First patch:
> 
> Helper functions and minor refactoring. 
> 
> Added explicit const to unchanged record reference for API clarity. Added helper functions for the reading and setting of fast-math flags and call arguments. Cleaned up the include lines.
> 
> 
> Second Patch: 
> 
> Intrinsic support for fast-math flags.
> 
> Previously, IntrinsicInsts were handled as CallInsts. However, intrinsics can now have fast-math flags, and thus merit a separate bitcode encoding. This change adds FUNC_CODE_INST_INTRINSIC for intrinsic calls, separates them out from general calls, and adds encoding and decoding of fast-math flags. Backwards compatibility is retained. Adds parsing fast-math flags for IntrinsicInsts, and more tests.

I'm pretty strongly against this.  A major feature of intrinsics (as opposed to well-known libcalls) is that they can have immediate operands to capture additional information.  For example, the known alignment of a memcpy is captured as an argument to intrinsic, even though the libc memcpy call doesn't have it.

Similarly, it doesn't make sense to make the concept of "intrinsics in general" more complicated to handle fast math.  "Fast math intrinsics" should either be modeled by adding new immediate operands to llvm.sin (and friends) or by using per-instruction metadata on these intrinsic calls.

-Chris