[llvm-bugs] [Bug 37269] New: [AArch64] Improve FNEG DAGCombine: (fneg (fmul c, x)) -> (fmul -c, x) on AArch64

[via llvm-bugs]

https://bugs.llvm.org/show_bug.cgi?id=37269

            Bug ID: 37269
           Summary: [AArch64]  Improve FNEG DAGCombine: (fneg (fmul c, x))
                    -> (fmul -c, x) on AArch64
           Product: libraries
           Version: trunk
          Hardware: PC
                OS: Linux
            Status: NEW
          Severity: enhancement
          Priority: P
         Component: Backend: AArch64
          Assignee: unassignedbugs at nondot.org
          Reporter: mcrosier at codeaurora.org
                CC: llvm-bugs at lists.llvm.org

The generic DAG combiner has the following rule in visitFNEG:

 (fneg (fmul c, x)) -> (fmul -c, x)

Given the following C code:

void t1(double a, double *ptr) {
  double b = 0.5 * a;
  ptr[0] = b;
  ptr[1] = -b;
}

Clang will generate the following assembly for AArch64 (with -O3
-ffp-contract=fast):
t1:
        fmov    d1, #0.50000000
        fmov    d2, #-0.50000000
        fmul    d1, d0, d1
        fmul    d0, d0, d2
        stp     d1, d0, [x0]
        ret

I can think of two alternative code sequences.

Sequence 1:
t1:
        fmov    d1, #0.50000000
        fmul    d2, d0, d1
        fnmul   d0, d0, d1
        stp     d2, d0, [x0]
        ret

Sequence 2:
t1:
        fmov    d1, #0.50000000
        fmul    d0, d0, d1
        fneg    d1, d0
        stp     d0, d1, [x0]

If we disable the aforementioned DAGCombine we'll get the first sequence.  I
would expect this sequence to be strictly better, since an extra fmov is
removed (which is a result of the first fmul having multipled uses).  GCC will
generate the code in sequence 2, but I haven't yet convinced myself this
sequence is better.  Specifically, while the fneg may have a lower latency as
compared to a fmul or fnmul (at least this is the case on Falkor), the fneg
must wait for the first fmul to complete, thus I expect the critical path to be
longer.

FWIW, I wanted to record my findings, but I'm not currently pursuing this
issue.

-- 
You are receiving this mail because:
You are on the CC list for the bug.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-bugs/attachments/20180427/971c9682/attachment.html>