[PATCH] D74851: [x86] use instruction-level fast-math-flags to drive MachineCombiner

[Michael Berg via Phabricator via llvm-commits]

mcberg2017 added a comment.

All in all LGTM, will check back in bit to see what others have to say.



================
Comment at: llvm/lib/Target/X86/X86InstrInfo.cpp:7661
+    return Inst.getFlag(MachineInstr::MIFlag::FmReassoc) &&
+           Inst.getFlag(MachineInstr::MIFlag::FmNsz);
   default:
----------------
The reassociate pass uses nsz too, and as a convention this seems fine.


================
Comment at: llvm/lib/Target/X86/X86InstrInfo.cpp:7850
+  // TODO: There should be a helper function for copying only fast-math-flags.
+  uint16_t IntersectedFlags = OldMI1.getFlags() & OldMI2.getFlags();
+  NewMI1.setFlags(IntersectedFlags);
----------------
I like this as for us in the case when we have divergent FMF environments brought together via inlining, this preserves the bounds of expression interfacing, where neither model would incur on the other.  Perhaps we should have a way to zap the int/float flags as a utility so that we can keep contexts separate in MachineInstr.h.


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D74851/new/

https://reviews.llvm.org/D74851