https://bugs.llvm.org/show_bug.cgi?id=37144

            Bug ID: 37144
           Summary: Codegen creates an extra bench in a binary search
           Product: new-bugs
           Version: unspecified
          Hardware: PC
                OS: Linux
            Status: NEW
          Severity: enhancement
          Priority: P
         Component: new bugs
          Assignee: unassignedb...@nondot.org
          Reporter: raf...@espindo.la
                CC: llvm-bugs@lists.llvm.org, r...@google.com

lld has a binary search function written as:

--------------------------------------------------
  while (Size != 1) {
    size_t H = Size / 2;
    const It MI = First + H;
    Size -= H;
    First = Comp(Value, *MI) ? First : First + H;
  }
  return Comp(Value, *First) ? First : First + 1;
--------------------------------------------------

The objective is to avoid branch misprediction. The body has no branches and
the loop has no early exits, so it is very predictable.

Everything is fine in the IR. The loop is a single BB and a select is used:

%12 = phi i32* [ %0, %8 ], [ %19, %10 ]
...
%14 = getelementptr inbounds i32, i32* %12, i64 %13
...
%18 = icmp ugt i64 %17, %2
%19 = select i1 %18, i32* %12, i32* %14

But codegen decides to use branches:

        cmpq    %rdx, %rcx
        ja      .LBB81_4
# %bb.3:                                #   in Loop: Header=BB81_2 Depth=1
        leaq    (%rdi,%rax,4), %rdi
.LBB81_4:                               #   in Loop: Header=BB81_2 Depth=1
        cmpq    $1, %rsi
        jne     .LBB81_2


Which causes massive amounts of branch misprediction. LLD's largest allocation
is a hash table that exists just to avoid using the binary search as often
since it is so expensive.

GCC gets this right. In a lld version without the extra hash table it looks
like fixing this bug would result in a 12% speedup of the entire link is some
testcases.

-- 
You are receiving this mail because:
You are on the CC list for the bug.
_______________________________________________
llvm-bugs mailing list
llvm-bugs@lists.llvm.org
http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs

Reply via email to