Re: [Design notes, RFC] Address-lowering prototype design (PR46556)

[Richard Guenther <richard.guenther@gmail.com>]

On Tue, Jun 14, 2011 at 4:18 PM, William J. Schmidt
<wschmidt@linux.vnet.ibm.com> wrote:
> On Tue, 2011-06-14 at 15:39 +0200, Richard Guenther wrote:
>> On Fri, Jun 10, 2011 at 5:11 PM, William J. Schmidt
>> <wschmidt@linux.vnet.ibm.com> wrote:
>> > On Tue, 2011-06-07 at 16:49 +0200, Richard Guenther wrote:
>> >> On Tue, Jun 7, 2011 at 4:14 PM, William J. Schmidt
>> >> <wschmidt@linux.vnet.ibm.com> wrote:
>> >
>> > <snip>
>> >
>> >> >> > Loss of aliasing information
>> >> >> > ============================
>> >> >> > The most serious problem I've run into is degraded performance due to poorer
>> >> >> > instruction scheduling choices.  I tracked this down to
>> >> >> > alias.c:nonoverlapping_component_refs_p.
>> >> >> >
>> >> >> > This code proves that two memory accesses don't overlap by attempting to prove
>> >> >> > that they access different fields of the same structure.  This is done using
>> >> >> > the MEM_EXPRs of the two rtx's, which record the expression trees that were
>> >> >> > translated into the rtx's during expand.  When address lowering is not
>> >> >> > present, a simple COMPONENT_REF will appear in the MEM_EXPR:  x.a, for
>> >> >> > example.  However, address lowering changes the simple COMPONENT_REF into a
>> >> >> > [TARGET_]MEM_REF that is no longer necessarily identifiable as a field
>> >> >> > reference.  Thus the aliasing machinery can no longer prove that two such
>> >> >> > field references are disjoint.
>> >> >> >
>> >> >> > This has severe consequences for performance, and has to be dealt with if
>> >> >> > address lowering is to be successful.
>> >> >> >
>> >> >> > I've worked around this with an admittedly fragile solution; I'll discuss the
>> >> >> > drawbacks below.  The idea is to construct a mapping from replacement mem_refs
>> >> >> > to the original expressions that they replaced.  When a MEM_EXPR is being set
>> >> >> > during expand, we first look up the mem_ref in the mapping.  If present, the
>> >> >> > MEM_EXPR is set to the original expression, rather than to the mem_ref.  This
>> >> >> > essentially duplicates the behavior in the absence of address lowering.
>> >> >>
>> >> >> Ick.  We had this in the past via TMR_ORIGINAL which caused all sorts
>> >> >> of problems.  Removing it didn't cause much degradation because we now
>> >> >> preserve points-to information.
>> >> >>
>> >> >> Originally I played with lowering all memory accesses to MEM_REFs
>> >> >> (see the old mem-ref branch), and the loss of type-based alias
>> >> >> disambiguation was indeed an issue.
>> >> >>
>> >> >> But - I definitely do not like the idea of preserving something similar
>> >> >> to TMR_ORIGINAL.  Instead we can try preserving some information
>> >> >> we derive from it.  We keep the original access type that we can use
>> >> >> for TBAA but do not retain knowledge on whether the type of the
>> >> >> MEM_REF is valid for TBAA or if it is view-converted.
>> >> >
>> >> > Yes, I really don't like what I have at the moment, either.  I put it in
>> >> > place as a stopgap to let me proceed to look for other performance
>> >> > problems.
>> >> >
>> >> > The question is how we can infer useful information for TBAA from the
>> >> > MEM_REFs and TMRs.  I poked at trying to identify types and offsets from
>> >> > the MEM_EXPRs, but this ended up being useless; I had to constrain too
>> >> > many cases to maintain correctness, and couldn't prove the type
>> >> > information for the important cases in SPEC I was trying to address.
>> >> >
>> >> > Unfortunately, the whole design goes down the drain if we can't find a
>> >> > way to solve the TBAA issue.  The performance degradations are too
>> >> > costly.
>> >>
>> >> If you look at what basic TBAA the alias oracle performs then it boils
>> >> down to the fact that get_alias_set for a.b.c might end up using the
>> >> alias-set of the type of C but for MEM[&a + 4] it will use the alias set
>> >> of the type of a.  The tree alias-oracle extracts both alias sets, that
>> >> of the outermost valid type and that of the innermost as both are
>> >> equally useful.  But the MEM_REF (or TARGET_MEM_REF) tree
>> >> only have storage for one such alias-set.  Thus my idea at some point
>> >> was to store the other one as well in some form.  It will not be
>> >> the full information (after all, the complete access path does provide
>> >> some extra information - see aliasing_component_refs_p).
>> >
>> > This is what concerns me.  TBAA information for the outer and inner
>> > components doesn't seem sufficient to provide what
>> > nonoverlapping_component_refs_p is currently able to prove.  The latter
>> > searches for a common RECORD_TYPE somewhere along the two access paths,
>> > and then disambiguates if the two associated referenced fields differ.
>> > For a simple case like "struct x { int a; int b; };", a and b have the
>> > same type and alias-set, so the alias-set information doesn't add
>> > anything.  It isn't sufficient alone for the disambiguation of x1.a =
>> > MEM_REF[&x1, 0] and x2.b = MEM_REF[&x2, 4].
>> >
>> > Obviously the offset is sufficient to disambiguate for this simple case
>> > with a common base type, but when the shared record types aren't at the
>> > outermost level, we can't detect whether it is.
>> >
>> > At the moment I don't see how we can avoid degradation unless we keep
>> > the full access path around somewhere, for [TARGET_]MEM_REFs built from
>> > COMPONENT_REFs.  I hope I'm wrong.
>>
>> You are not wrong.  But the question is, does it make a difference?
>>
>> Richard.
>
> Yes, it does.  This scenario occurs in 188.ammp, among others, and leads
> to a large degradation without the change.  The performance-critical
> loop in mm_fv_update_nonbon makes heavy use of indirect references to
> the ATOM structure that contains numerous float variables.  When the
> COMPONENT_REFs have been converted to MEM_REFs, the alias machinery can
> no longer disambiguate these, which constrains the scheduler.  The
> result of the poor scheduling (on PowerPC, at least) is a large increase
> in floating-point spill code in the critical loop.

As they appear in loops I wonder why IVOPTs doesn't already expose
this problem?

In general the answer to missed TBAA optimizations is of course
"make sure that PTA works", which usually means using LTO ...

I really really don't like preserving TBAA related information as trees.
Instead if we really think preserving access paths is a must (I have
already significantly reduced the number of preserved paths with
introducing MEM_REFs!) then we have to find a different representation.

I suppose you can turn the AMMP case into a (small) testcase that
illustrates the issue and allows easier analysis and test of fixes?

Richard.