Flag output used by two other nodes in DAG

Project Infrastructure LLVM Dev List Archives
1

I recently filed this bug: http://llvm.org/bugs/show_bug.cgi?id=8323
It's a dodgy one because you have to patch LLVM to demonstrate it.

I suspect that the cause of the problem in that "bug" is that the
peephole optimisation in PerformDAGCombine results in a Flag output
from one node being used as input by two other nodes in the DAG, and
the scheduler then can't cope with that.

Is it, or should it be legal for a Flag output to be used as input by
more than one other node?

If it is legal, does it ever actually work, and, if not, is there
already a bug filed relating to this?

If it isn't legal, are there sufficient guarantees that it won't
happen, and could we please have a better error message if it does
happen by accident?

In either case, how should the following code get translated in such a
way that we don't repeat the comparison?

int f1(int x)
{
  return x < 0 ? 11 : x == 0 ? 22 : 33;
}

Currently, with the ARM back end, I'm getting "cmp r0, #0" repeated.

Replace 0 by 100 and I get two instances of "cmpl $100, %edi" in the
x86 output, too.

Is there already a bug filed relating to that missed optimisation?

(Bug 7592 is related to it.)

Thanks,

Edmund

2

Hello, Edmund,

Is it, or should it be legal for a Flag output to be used as input by
more than one other node?

It's illegal. Multiple uses of the flag output do not make any sense,
this breaks the semantics of flag operands.

3

Anton Korobeynikov:

> Is it, or should it be legal for a Flag output to be used as input by
> more than one other node?
It's illegal. Multiple uses of the flag output do not make any sense,
this breaks the semantics of flag operands.

All right. Then what should the Selection DAG look like in a case where the
flag value generated by one instruction is to be used as input to two other
instructions?

For a concrete example, consider:

int f(int x)
{
  return x < 0 ? 11 : x == 0 ? 22 : 33;
}

I'd like this to turn into something like what I've seen from other compilers:

  cmp r0, #0
  movlt r0, #11
  bxlt lr
  movne r0, #33
  moveq r0, #22
  bx lr

4

All right. Then what should the Selection DAG look like in a case where the
flag value generated by one instruction is to be used as input to two other
instructions?

It is impossible. You cannot "glue" two nodes to one. So, you need to
build a sequence of flag use / flag def nodes.