[PATCH] D70226: Add an option to disable strict float node mutating to an normal float node

[Ulrich Weigand via Phabricator via llvm-commits]

uweigand added a comment.

In D70226#1752432 <https://reviews.llvm.org/D70226#1752432>, @craig.topper wrote:

> I'm strongly considering making mutation a true operation action like Expand, Legal, Custom, etc. So we can distinquish Expand from "my target doesn't support strict FP yet". The checks for "legal" on the non-strict nodes to guess what we should do, don't work for X86 where we have a lot of Custom handling. The strict fp operations would default to this new operation action instead of Expand.


Hmm.  I'd actually consider this a step in the wrong direction, since the mutation operation is really wrong, it doesn't actually respect strict fp semantics.  So I'd rather have it go away completely as soon as possible -- once we have enough target coverage (e.g. X86, Arm, Power, SystemZ?).  That's why I like the approach in this patch: all the mutation-related stuff is now behind the isStrictFPEnabledCheck, and soon as we decide target support is sufficient, we just remove all that code.  (For targets that still don't support strict FP natively, the intrinsics would then generally map to libcalls.)


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D70226/new/

https://reviews.llvm.org/D70226