go/types: sort to reduce computational complexity of initOrder
Our calculation of initOrder builds the dependency graph and then
removes function nodes approximately at random. While profiling, I
noticed that this latter step introduces a superlinear algorithm into
our type checking pass, which can dominate type checking for large
packages such as runtime.
It is hard to analyze this rigorously, but to give an idea of how such a
non-linearity could arise, suppose the following assumptions hold:
- Every function makes D calls at random to other functions in the
package, for some fixed constant D.
- The number of functions is proportional to N, the size of the package.
Under these simplified assumptions, the cost of removing an arbitrary
function F is P*D, where P is the expected number of functions calling
F. P has a Poisson distribution with mean D.
Now consider the fact that when removing a function F in position i, we
recursively pay the cost of copying F's predecessors and successors for
each node in the remaining unremoved subgraph of functions containing F.
With our assumptions, the size of this subgraph is proportional to
(N-i), the number of remaining functions to remove.
Therefore, the total cost of removing functions is proportional to
P*D*Σᴺ(N-i)
which is proportional to N².
However, if we remove functions in ascending order of cost, we can
partition by the number of predecessors, and the total cost of removing
functions is proportional to
N*D*Σ(PMF(X))
where PMF is the probability mass function of P. In other words cost is
proportional to N.
Assuming the above analysis is correct, it is still the case that the
initial assumptions are naive. Many large packages are more accurately
characterized as combinations of many smaller packages. Nevertheless, it
is intuitively clear that removing expensive nodes last should be
cheaper.
Therefore, we sort by cost first before removing nodes in
dependencyGraph.
We also move deletes to the outer loop, to avoid redundant deletes. By
inspection, this avoids a bug where n may not have been removed from its
successors if n had no predecessors.
name old time/op new time/op delta
Check/runtime/funcbodies/noinfo-8 568ms ±25% 82ms ± 1% -85.53% (p=0.000 n=8+10)
name old lines/s new lines/s delta
Check/runtime/funcbodies/noinfo-8 93.1k ±56% 705.1k ± 1% +657.63% (p=0.000 n=10+10)
Updates #49856
Change-Id: Id2e70d67401af19205e1e0b9947baa16dd6506f0
Reviewed-on: https://go-review.googlesource.com/c/go/+/369434
Trust: Robert Findley <rfindley@google.com>
Run-TryBot: Robert Findley <rfindley@google.com> Reviewed-by: Robert Griesemer <gri@golang.org>
TryBot-Result: Gopher Robot <gobot@golang.org>