[llvm-dev] [RFC] NewGVN

Daniel Berlin via llvm-dev llvm-dev at lists.llvm.org
Tue Nov 15 23:29:41 PST 2016

First, thanks. This is a very very long time coming :)
Second, for those watching, note that pretty much all of the improvements
and missing cases, including load forwarding/coercion, etc, actually are
It's more a matter of cleaning them up, breaking it down, and submitting
it, than "implementing them".
The main experimentation at this point is "can we do it cleaner" not "can
we do it" :)

It's also important to note that this new GVN also treats loads/stores in a
unified way with scalars, unlike current GVN (which has no load or store
value numbering).

So it will happily discover complex load/store relations (though there is
some improvements we can still make here)
For example:

int vnum_test8(int *data)
  int i;
  int stop = data[3];
  int m = data[4];
  int n = m;
  int p;
  for (i=0; i<stop; i++) {
    int k = data[2];
    data[k] = 2;
    data[0] = m - n;
    k = data[1];
    m = m + k;
    n = n + k;
    p = data[0];
  return p;

LLVM's current GVN will eliminate a single load here[1]
NewGVN will calculate that m and n are equal, that m-n is 0, that p is 0

It's not quite perfect yet, i haven't fixed store handling, so the
following is missed:
int a;
int *p;
// LLVM is too smart and if we don't do this, realizes *p is a store to
void foo(){
    p = &a;
int main(int argc, char **argv) {
  int result;
  *p = 2;
  if (argc)
    *p = 2;
  result = *p;
  return result;

Here, current LLVM GVN will do nothing, because it can't understand
anything really about the stores.
GCC's GVN will determine result is 2.
NewGVN is not quite that smart yet (it require a little work to what we do
to stores, and value numbering memory ssa versions)

This issue compounds if you have conditional stores of the same value.

So, for example, if you add:

if (i < 30)
  data[0] = 0;

to the first case.

GCC can still determine p is 0.

Currently, NewGVN cannot.


On Tue, Nov 15, 2016 at 10:49 PM, Davide Italiano <davide at freebsd.org>

> Hi,
> we would like to propose  a new Global Value Numbering pass in LLVM.
> The ideas/code are from Daniel Berlin (with a minor overhaul/splitting
> into submittable patches from me). The code has been around for a
> while (2012 or before), and we think it's getting ready to be
> committed upstream.
> ### Motivation
> To put things into context: my personal motivation for having a new
> GVN/PRE algorithm is LTO.
> It's not a secret that LLVM is getting slower and slower release after
> release, as Rafael discovered/pointed out in March [1] (and probably
> many others found out). I personally took a shot at profiling LTO on
> many internal/opensource applications (including clang itself) and
> noticed that GVN always show in the top-3 passes (and it's generally
> the pass where we spend most of the time in the middle-end). There are
> cases (extreme) where 90% of the compile time goes in GVN.
> Example:
> ===---------------------------------------------------------
> ----------------===
>                       ... Pass execution timing report ...
> ===---------------------------------------------------------
> ----------------===
>   Total Execution Time: 684.7622 seconds (683.7141 wall clock)
>    ---User Time---   --System Time--   --User+System--   ---Wall
> Time---  --- Name ---
>   130.2558 ( 20.0%)   6.3128 ( 18.7%)  136.5686 ( 19.9%)  137.6145 (
> 20.1%)  X86 DAG->DAG Instruction Selection
>   55.4695 (  8.5%)   0.5501 (  1.6%)  56.0196 (  8.2%)  56.1049 (
> 8.2%)  Function Integration/Inlining
>   42.3492 (  6.5%)   0.0364 (  0.1%)  42.3856 (  6.2%)  42.8676 (
> 6.3%)  Global Value Numbering
> ### Problems in the current GVN
> There are some issues in the current GVN infrastructure. A
> non-exhaustive list (I'm pretty sure people have their list of
> complains, these are the ones I care about the most).
> * GVN is very slow for large test cases
> * GVN doesn't do a real analysis, instead it eliminates as it goes, so
> it's hard to reason about it.
> * GVN doesn't perform any phi predication, i.e. doesn't know about phi
> nodes, so later passes have to do some extra work to clean up
> * There are bugs, e.g. [2] which would require a rewrite of PRE to be
> fixed.
> ### NewGVN
> The new algorithm implements the ideas described in [3] with some
> engineering optimizations by Dan (for example the set of touched
> instructions is represented using a Bitvector instead of a set because
> it's not uncommon for large functions where a predicate change that
> thousands of instructions need to be changed, and we both measured 30%
> of the whole pass time spent just modifying the set).
> The code pretty much maps what the paper describe so I won't try to
> repeat it here =)
> Some advantages of NewGVN:
> * GVN performs a real analysis (which is now part of the pass itself
> but potentially could be split and used as an utility by other
> passes). For example, Quentin/Dan pointed out that outlining at the IR
> level is hard without a proper value numbering analysis.
> * On large testcases, It's faster than current GVN, even though didn't
> go through a lot of profiling/optimization lately (I found it up to
> 50% faster in compile time on some internal benchmarks). There are
> some places were we can improve. For example, we spend a lot of time
> inside Simplify* functions. Another considerable chunk of the time is
> spent inside MemorySSA, but this is going to change once we preserve
> MemSSA more aggressively.
> * The code is easier to understand (at least to me)
> [...]
> Some current limitations of NewGVN:
> * NewGVN doesn't do everything that the current GVN does. There are
> plans to implement missing features and more.
> * On small testcases NewGVN may be slower as it does some work upfront
> that the current GVN doesn't do. NewGVN is probably less lazy than it
> should be, but it didn't matter for us so we didn't consider it a
> priority. If somebody cares and finds a case where NewGVN
> substantially regresses, speak up.
> The initial code can be found here https://reviews.llvm.org/D26224
> The current patch includes only the "core" GVN algorithm, i.e. the
> expression framework/the logic to perform congruence finding. Other
> pieces (e.g. PRE will build on top of that).
> My rough plan is:
> * Try to get this reviewed/tested and in tree.
> * Fix miscompiles/bugs (and performance problems as/if they arise).
> * Build other pieces of the algorithm on top of what we have. My
> immediate concern will be implementing support for llvm.assume and
> load coercion. Dan said he will try to find the time to work on the
> predicate handling (actually, there's already a branch
> https://github.com/dberlin/llvm-gvn-rewrite/tree/newgvn-predicates).
> Please let us know what you think. Any feedback/review comment/testing
> is very appreciated!
> Thanks!
> [1] http://lists.llvm.org/pipermail/llvm-dev/2016-March/096488.html
> [2] https://llvm.org/bugs/show_bug.cgi?id=30692#c11
> [3] http://dl.acm.org/citation.cfm?id=512536
> --
> Davide
> "There are no solved problems; there are only problems that are more
> or less solved" -- Henri Poincare
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20161115/276e16ae/attachment.html>

More information about the llvm-dev mailing list