Bug ID: 36357
           Summary: [x86] Revert r319777(clang) and r319778(llvm) and
                    r319911(llvm) from 6.0 branch
           Product: libraries
           Version: trunk
          Hardware: PC
                OS: Windows NT
            Status: NEW
          Severity: enhancement
          Priority: P
         Component: Backend: X86

These commits removed some intrinsics and implemented them with native IR. But
it turns out that we may have messed up the operand order in the upgrade due to
inconsistent documentation from Intel.

Intel documentation says this for kunpackb intrinsic

  k[7:0] := b[7:0]
  k[15:8] := a[7:0]
  k[MAX:16] := 0

But for kunpackw it says

  k[31:0] := a[31:0]
  k[63:32] := b[31:0]
  k[MAX:64] := 0

Notice the order of 'a' and 'b' is reversed here. kunpackd is documented
similarly to kunpackw.

It turns out icc, gcc, and clang 5.0 all implement the order specified for
kunpackb for all 3 intrinsics.

It seems that when the upgrade was done in 6.0, we followed the
kunpackw/kunpackd documentation instead.

There have been other changes to the autoupgrade here after the 6.0 branch. So
rather than try to fix the 6.0 implementation and the trunk implementation
separately, I propose that we revert the commits on the 6.0 branch.

You are receiving this mail because:
You are on the CC list for the bug.
llvm-bugs mailing list

Reply via email to