std::distance(C->Relocs.end(), C->Relocs.begin()) is the same as NumRelocs
which is already added to the hash value. What we are missing here is the
section size.
llvm-svn: 248202
This patch fixes a regression introduced by r247964. Relocations that
are referring the same symbol should be considered equal, but they
were not if they were pointing to non-section chunks.
llvm-svn: 248132
The LLD's ICF algorithm is highly parallelizable. This patch does that
using parallel_for_each.
ICF accounted for about one third of total execution time. Previously,
it took 324 ms when self-hosting. Now it takes only 62 ms.
Of course your mileage may vary. My machine is a beefy 24-core Xeon machine,
so you may not see this much speedup. But this optimization should be
effective even for 2-core machine, since I saw speedup (324 ms -> 189 ms)
when setting parallelism parameter to 2.
llvm-svn: 248038
Previously, ICF created a vector for each SectionChunk. The vector
contained pointers to successors, which are namely associative sections
and COMDAT relocation targets. The reason I created vectors is because
I thought that that would make section comparison faster.
It did make the comparison faster. When self-linking, for example, it
saved about 10 ms on each iteration. The time we spent on constructing
the vectors was 124 ms. If we iterate more than 12 times, return from
the investment exceeds the initial cost.
In reality, it usually needs 5 iterations. So we shouldn't construct
the vectors.
llvm-svn: 247963
equalsConstants() is the heaviest function in ICF, and that consumes
more than half of total ICF execution time. Of which, section content
comparison accounts for roughly one third.
Previously, we compared section contents at the beginning of the
function after comparing their checksums. The comparison is very
likely to succeed because when the control reaches that comparison,
their checksums are always equal. And because checksums are 64-bit
CRC, they are unlikely to collide.
We compared relocations and associative sections after that.
If they are different, the time we spent on byte-by-byte comparison
of section contents were wasted.
This patch moves the comparison at the end of function. If the
comparison fails, the time we spent on relocation comparison are
wasted, but as I wrote it's very unlikely to happen.
LLD took 1198 ms to link itself to produce a 27.11 MB executable.
Of which, ICF accounted for 536 ms. This patch cuts it by 90 ms,
which is 17% speedup of ICF and 7.5% speedup overall. All numbers
are median of ten runs.
llvm-svn: 247961
This patch defines ICF class and defines ICF-related functions as
members of the class. By doing this we can move code that are
related only to ICF from SectionChunk to the newly-defined class.
This also eliminates a global variable "NextID".
llvm-svn: 247802
This is a patch to make LLD to be on par with MSVC in terms of ICF
effectiveness. MSVC produces a 27.14MB executable when linking LLD.
LLD previously produced a 27.61MB when self-linking. Now the size
is reduced to 27.11MB. Note that without ICF the size is 29.63MB.
In r247387, I implemented an algorithm that handles section graphs
as cyclic graphs and merge them using SCC. The algorithm did not
always work as intended as I demonstrated in r247721. The new
algortihm implemented in this patch is different from the previous
one. If you are interested the details, you want to read the file
comment of ICF.cpp.
llvm-svn: 247770
Previously, LLD's ICF couldn't merge cyclic graphs. That was unfortunate
because, in COFF, cyclic graphs are not exceptional at all. That is
pretty common.
In this patch, sections are grouped by Tarjan's strongly connected
component algorithm to get acyclic graphs. And then we try to merge
SCCs whose outdegree is zero, and remove them from the graph. This
makes other SCCs to have outdegree zero, so we can repeat the
process until all SCCs are removed. When comparing two SCCs, we handle
cycles properly.
This algorithm works better than previous one. Previously, self-linking
produced a 29.0MB executable. It now produces a 27.7MB. There's still some
gap compared to MSVC linker which produces a 27.1MB executable for the
same input. So the gap is narrowed, but still LLD is not on par with MSVC.
I'll investigate that later.
llvm-svn: 247387
Identical COMDAT Folding is a feature to merge COMDAT sections
by contents. Two sections are considered the same if their contents,
relocations, attributes, etc, are all the same.
An interesting fact is that MSVC linker takes "iterations" parameter
for ICF because the algorithm they are using is iterative. Merging
two sections could make more sections to be mergeable because
different relocations could now point to the same section. ICF is
repeated until we get a convergence (until no section can be merged).
This algorithm is not fast. Usually it needs three iterations until a
convergence is obtained.
In the new algorithm implemented in this patch, we consider sections
and relocations as a directed acyclic graph, and we try to merge
sections whose outdegree is zero. Sections with outdegree zero are then
removed from the graph, which makes other sections to have outdegree
zero. We repeat that until all sections are processed. In this
algorithm, we don't iterate over the same sections many times.
There's an apparent issue in the algorithm -- the section graph is
not guaranteed to be acyclic. It's actually pretty often cyclic.
So this algorithm cannot eliminate all possible duplicates.
That's OK for now because the previous algorithm was not able to
eliminate cycles too. I'll address the issue in a follow-up patch.
llvm-svn: 246878
Previously, we calculated our own hash values for section contents.
Of coruse that's slow because we had to access all bytes in sections.
Fortunately, COFF objects usually contain hash values for COMDAT
sections. We can use that to speed up Identical COMDAT Folding.
llvm-svn: 246869
Only SectionChunk can be dead-stripped. Previously,
all types of chunks implemented these functions,
but their functions were blank.
Likewise, only DefinedRegular and DefinedCOMDAT symbols
can be dead-stripped. markLive() function was implemented
for other symbol types, but they were blank.
I started thinking that the change I made in r240319 was
a mistake. I separated DefinedCOMDAT from DefinedRegular
because I thought that would make the code cleaner, but now
we want to handle them as the same type here. Maybe we
should roll it back.
This change should improve readability a bit as this removes
some dubious uses of reinterpret_cast. Previously, we
assumed that all COMDAT chunks are actually SectionChunks,
which was not very obvious.
llvm-svn: 240675
Rui Ueyama