[LLVMdev] Load serialisation during selection DAG building

[Hal Finkel]

Steve,

I had created a patch last year that does something similar to what you
describe for regular loads and stores using aliasing information. I
think that the last message in the thread was:
http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20120402/140299.html

This approach has worked for me, but it is not the preferred solution
going forward. The preferred solution is to keep the "critical
chain" as is (partly because there are places in some of the backend
lowering code that assume it exists in its current form), and just
update the scheduler to be more intelligent about how it forms the
dependency graph. This has since been developed for the new MI
scheduling framework by Sergei Larin, and was committed in r156842
(last message in the discussion thread was
http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20120507/142659.html)

I would recommend trying something like Sergei's solution first, and
fall back to trying to play with the critical chain only if that can't
or won't work.

 -Hal

On Mon, 13 Aug 2012 11:29:18 +0100
Steve Montgomery <stephen.montgomery3 at btinternet.com> wrote:

> I've got a question about how SelectionDAGBuilder treats loads.
> 
> The LLVM Language Reference Manual explicitly states that the order
> of volatile operations may be changed relative to non-volatile
> operations. However, when the SelectionDAGBuilder in LLVM 3.1
> encounters a volatile load, it flushes all pending loads and then
> chains the volatile load onto them meaning that the volatile load
> must be scheduled after those loads. While this behaviour isn't
> wrong, it seems to reduce the scope for efficient instruction
> selection.
> 
> Is there any reason not to modify the behaviour of
> SelectionDAGBuilder::visitLoad() so that volatile loads don't cause
> SelectionDAGBuilder::getRoot() to be called. Instead, they can be
> chained together with the head of the chain being stored in
> PendingLoads. Then when something else calls
> SelectionDAGBuilder::getRoot(), the chain of volatile loads is
> TokenFactored together with the non-volatile loads. I've tried this
> out and it seems to do what I want but as I'm fairly inexperienced
> with LLVM, I'm not sure whether there's something else preventing
> this strategy from working.
> 
> The reason I noticed this is because I have been developing a
> back-end for a target in which some instructions are implemented as
> pseudos which will be expanded into a pair of instructions. Each of
> the two operands of the pseudo can be a load but as the expansion
> accesses the right operand twice, I don't want to match a volatile
> load for this operand. For example, in:
> 
>   %0 = load i16* @y, align 2, !tbaa !0
>   %1 = load volatile i16* @x, align 2, !tbaa !0
>   %add = add i16 %0, %1 
> 
> the volatile load is sequenced after the non-volatile load which
> means that the non-volatile load can't match the left operand of the
> add because this would create a scheduling cycle. This means both
> loads are selected as load instructions, resulting in use of an extra
> register and an extra instruction. If I change the behaviour of
> SelectionDAGBuilder::visitLoad() as described then I get just two
> instructions. _______________________________________________ LLVM
> Developers mailing list LLVMdev at cs.uiuc.edu
> http://llvm.cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev



-- 
Hal Finkel
Postdoctoral Appointee
Leadership Computing Facility
Argonne National Laboratory