replacing the backwards threader and more

[Aldy Hernandez]

On 6/28/21 10:22 AM, Richard Biener wrote:
> On Fri, Jun 25, 2021 at 6:20 PM Aldy Hernandez <aldyh@redhat.com> wrote:
>>
>> Hi folks.
>>
>> I'm done with benchmarking, testing and cleanups, so I'd like to post my
>> patchset for review.  However, before doing so, I'd like to address a
>> handful of meta-issues that may affect how I post these patches.
> 
> First of all thanks for the detailed analysis and report!
> 
>> Trapping on differences
>> =======================
>>
>> Originally I wanted to contribute verification code that would trap if
>> the legacy code threaded any edges the new code couldn't (to be removed
>> after a week).  However, after having tested on various architectures
>> and only running once into a missing thread, I'm leaning towards
>> omitting the verification code, since it's fragile, time consuming, and
>> quite hacky.
> 
> Agreed.
> 
>> For the record, I have tested on x86-64, aarch64, ppc64 and ppc64le.
>> There is only one case, across bootstrap and regression tests where the
>> verification code is ever tripped (discussed below).
>>
>> Performance
>> ===========
>>
>> I re-ran benchmarks as per our callgrind suite, and the penalty with the
>> current pipeline is 1.55% of overall compilation time.  As is being
>> discussed, we should be able to mitigate this significantly by removing
>> other threading passes.
> 
> 1.55% for the overall compilation looks quite large - is this suite particularly
> demanding on VRP/ranger?  Is the figure for 'cc1files' similar? (collect
> preprocessed sources of gcc/*.{c,cc} compiles, like by appending
> -save-temps to CFLAGS in a non-bootstrap build)

The figures are for cc1 files.  They are .ii files we run under 
callgrind and compare "instruction counts" as per the tool.  We've found 
it's comparable to -ftime-report, though not exactly.  We've been using 
this for benchmarking, since evrp times were so small that it was hard 
to compare user times.

> 
> I'm curious if you tried to look at the worst offenders and if there's
> anything new, threading specific.  More threading might result in
> larger code (how is cc1 size affected by the change?) and more
> code to compile by later passes can easily explain >1% differences.

A while ago, I did look at worst offenders on the branch, but there was 
nothing obvious.  As it stands now, the average cost for the threading 
pass is 858%, with 95% of .ii files falling under 1800%.  I could look 
at the outliers which seem to be c-fold, gcc-ar, gcc-ranlib, gcc-nm, and 
c-opts.

cc1 on x86-64 seems to be 0.135% larger for an --enable-checking build.

> 
>> Failing testcases
>> =================
>>
>> I have yet to run into incorrect code being generated, but I have had to
>> tweak a considerable number of tests.  I have verified every single
>> discrepancy and documented my changes in the testsuite when it merited
>> doing so.  However, there are a couple tests that trigger regressions
>> and I'd like to ask for guidance on how to address them.
>>
>> 1. gcc.c-torture/compile/pr83510.c
>>
>> I would like to XFAIL this.
>>
>> What happens here is that thread1 threads a switch statement such that
>> the various cases have been split into different independent blocks.
>> One of these blocks exposes an arr[i_27] access which is later
>> propagated by VRP to be arr[10].  This is an invalid access, but the
>> array bounds code doesn't know it is an unreachable path.
>>
>> However, it is not until dom2 that we "know" that the value of the
>> switch index is such that the path to arr[10] is unreachable.  For that
>> matter, it is not until dom3 that we remove the unreachable path.
> 
> Sounds reasonable.
> 
>> 2. -Wfree-nonheap-object
>>
>> This warning is triggered while cleaning up an auto_vec.  I see that the
>> va_heap::release() inline is wrapped with a pragma ignore
>> "-Wfree-nonheap-object", but this is not sufficient because jump
>> threading may alter uses in such a way that may_emit_free_warning() will
>> warn on the *inlined* location, thus bypassing the pragma.
>>
>> I worked around this with a mere:
>>
>>   > @@ -13839,6 +13839,7 @@ maybe_emit_free_warning (tree exp)
>>>     location_t loc = tree_inlined_location (exp);
>>> +  loc = EXPR_LOCATION (exp);
>>
>> but this causes a ton of Wfree-nonheap* tests to fail.  I think someone
>> more knowledgeable should address this (msebor??).
> 
> That looks wrong, but see msebors response for maybe some better answer.

Oh, it was just a stop-gap.  I have instead opted for the following 
temporary measure in Makefile.in:

tree-ssa-loop-im.o-warn = -Wno-free-nonheap-object

I will work with Martin on a more appropriate solution.

> 
>> 3. uninit-pred-9_b.c
>>
>> The uninit code is getting confused with the threading and the bogus
>> warning in line 24 is back.  I looked at the thread, and it is correct.
>>
>> I'm afraid all these warnings are quite fragile in the presence of more
>> aggressive optimizations, and I suspect it will only get worse.
> 
> Yep.  Shouldn't be a blocker, just XFAIL ...
> 
>> 4. libphobos/src/std/net/isemail.d
>>
>> This is a D test where we don't actually fail, but we trigger the
>> verification code.  It is the only jump threading edge that the new code
>> fails to get over the old code, and it only happens on ppc64.
>>
>> It triggers because a BB4 -> BB5 is too expensive to thread, but a BBn
>> -> BB3 -> BB4 -> BB5 is considered safe to thread because BB3 is a latch
>> and it alters the profitability equation.  The reason we don't get it,
>> is that we assume that if a X->Y is unprofitable, it is not worth
>> looking at W->X->Y and so forth.
>>
>> Jeff had some fancy ideas on how to attack this.  Once such idea was to
>> stop looking back, but only for things we were absolutely sure would
>> never yield a profitable path.  I tried a subset of this, by allowing
>> further looks on this latch test, but my 1.55% overall performance
>> penalty turned into an 8.33% penalty.  Personally it looks way too
>> expensive for this one isolated case.  Besides, the test where this
>> clamping code originally came from still succeeds (commit
>> eab2541b860c48203115ac6dca3284e982015d2c).
> 
> ick - I definitely expect some corner case optimization "regressions"
> caused by the change if and only because it might trigger pass
> ordering issues with later passes not expecting "optimized" code.
> So I'd disregard a single "missed" case for this moment.  It might
> be interesting to try to distill a C testcase from the case and file it
> with bugzilla for reference purposes.

Assuming there's nothing inherently D-ish in the testcase, I can 
certainly do this.

> 
>> CONCLUSION
>> ==========
>>
>> That's basically it.
>>
>> If we agree the above things are not big issues, or can be addressed as
>> follow-ups, I'd like to start the ball rolling on the new threader.
>> This would allow more extensive testing of the code, and separate it a
>> bit from the other big changes coming up :).
> 
> I'd say go ahead (with the bogus change somehow resolved), I'd still
> like to know sth about the compile-time issue but then we can eventually
> deal with this as followup as well.  If size is the reason we can see taming
> down the threader via some --param adjustments for example.
> 
> I'm just trying to make sure we're not introducing some algorithmic
> quadraticnesses that were not there before.


Agreed.  FWIW, there are already various --param knobs in the backwards 
threader we could tweak, though in practice the truly problematic cases 
we chop right off the bat, as soon as we go down an unprofitable path.

Thanks.
Aldy