Previously we were taking over 13 minutes to link Firefox's xul.dll
on ARM64; this reduces link time to around 18s on my machine.
The root cause of the problem was that all of the input .pdata sections
had the same unrelocated section data and therefore the same hash,
which made segregation quadratic in the number of .pdata sections. The
reason why we weren't observing this on other architectures was that
ARM has a different .pdata format. On non-ARM the format is (start
address, end address, .xdata), which caused the size of the function
to appear in the unrelocated section data where the end address field
is. However, the ARM format omits the end address field.
Fixes PR39667.
Differential Revision: https://reviews.llvm.org/D54809
llvm-svn: 347429
After fixing up the runtime pseudo relocation, the .refptr.<var>
will be a plain pointer with the same value as the IAT entry itself.
To save a little binary size and reduce the number of runtime pseudo
relocations, redirect references to the IAT entry (via the __imp_<var>
symbol) itself and discard the .refptr.<var> chunk (as long as the
same section chunk doesn't contain anything else than the single
pointer).
As there are now cases for both setting the Live variable to true
and false externally, remove the accessors and setters and just make
the variable public instead.
Differential Revision: https://reviews.llvm.org/D51456
llvm-svn: 341175
This patch does the same thing as r338153 for COFF.
Note that this patch affects only the order of log messages.
The output file is already deterministic.
Differential Revision: https://reviews.llvm.org/D50023
llvm-svn: 338406
This is needed to avoid merging two functions with identical
instructions but different xdata. It also reduces binary size by
deduplicating identical pdata sections.
Fixes PR35337.
Differential Revision: https://reviews.llvm.org/D46672
llvm-svn: 332169
In an upcoming change I will need to make a distinction between section
type (code, data, bss) and permissions. The term that I use for both
of these things is "output characteristics".
Differential Revision: https://reviews.llvm.org/D45799
llvm-svn: 330361
In COFF, duplicate string literals are merged by placing them in a
comdat whose leader symbol name contains a specific prefix followed
by the hash and partial contents of the string literal. This gives
us an easy way to identify sections containing string literals in
the linker: check for leader symbol names with the given prefix.
Any sections that are identified in this way as containing string
literals may be tail merged. We do so using the StringTableBuilder
class, which is also used to tail merge string literals in the ELF
linker. Tail merging is enabled only if ICF is enabled, as this
provides a signal as to whether the user cares about binary size.
Differential Revision: https://reviews.llvm.org/D44504
llvm-svn: 327668
Previously wasm used a separate header to declare markLive
and ELF used to declare ICF. This change makes each backend
consistently declare these in their own headers.
Differential Revision: https://reviews.llvm.org/D43529
llvm-svn: 325631
The classes used to print and update time information are in
common, so other linkers could use this as well if desired.
Differential Revision: https://reviews.llvm.org/D41915
llvm-svn: 322736
This requirement was added in r254578 to fix pr25686. However, it
appears to have originated from a misdiagnosis of the problem: link.exe
refused to merge the two sections because they are non-executable,
not because they have internal leaders. If I set up a similar scenario
with functions instead of globals I see that link.exe merges them.
Differential Revision: https://reviews.llvm.org/D40236
llvm-svn: 318682
Summary:
Many small functions have identical unwind info because they push the
same sets of CSRs in the same order and have the same stack and prologue
size. The VC linker merges duplicate .xdata, and so should LLD.
This reduces the .xdata section size of clang.exe from 1.8MB to 94KB.
Reviewers: pcc, ruiu
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D40160
llvm-svn: 318547
Now that we have only SymbolBody as the symbol class. So, "SymbolBody"
is a bit strange name now. This is a mechanical change generated by
perl -i -pe s/SymbolBody/Symbol/g $(git grep -l SymbolBody lld/ELF lld/COFF)
nd clang-format-diff.
Differential Revision: https://reviews.llvm.org/D39459
llvm-svn: 317370
Summary:
The COFF linker and the ELF linker have long had similar but separate
Error.h and Error.cpp files to implement error handling. This change
introduces new error handling code in Common/ErrorHandler.h, changes the
COFF and ELF linkers to use it, and removes the old, separate
implementations.
Reviewers: ruiu
Reviewed By: ruiu
Subscribers: smeenai, jyknight, emaste, sdardis, nemanjai, nhaehnle, mgorny, javed.absar, kbarton, fedor.sergeev, llvm-commits
Differential Revision: https://reviews.llvm.org/D39259
llvm-svn: 316624
Computing section content hashes early seems like a win in terms of
performance. It increases a chance that two different sections will get
different class IDs from the beginning.
Without threads, this patch improves Chromium link time by about 0.3
seconds. With threads, by 0.1 seconds. That's less than 1% time saving
but not bad for a small patch.
llvm-svn: 314644
This is the only place we use threads for ICF. The intention of this code
was to split an input vector into 256 shards and process them in parallel.
What the code was actually doing was to split an input into 257 shards,
process the first 256 shards in parallel, and the remaining one in serial.
That means this code takes ceil(256/n)+1 instead of ceil(256/n) where n
is the number of available CPU cores. The former converges to 2 while
the latter converges to 1.
This patches fixes the above issue.
llvm-svn: 303797
This seems to be the behavior of the MSVC linker. Previously, this
incompatibility caused nasty issues in chromium build a few times.
Differential Revision: https://reviews.llvm.org/D30363
llvm-svn: 301598
LLD is a multi-threaded program. errs() or outs() are not guaranteed
to be thread-safe (they are actually not).
LLD's message(), log() or error() are thread-safe. We should use them.
llvm-svn: 295787
This ports the ELF linker's symbol table design, introduced in r268178,
to the COFF linker.
Differential Revision: http://reviews.llvm.org/D21166
llvm-svn: 289280
LLD used to take 11.73 seconds to link Clang. Now it is 6.94 seconds.
MSVC link takes 83.02 seconds. Note that ICF is enabled by default on
Windows, so a low latency ICF is more important than in ELF.
llvm-svn: 288487
Associative sections are sections that need to be linked if their associated
sections are linked. Associative sections are used to append auxiliary data
such as debug info.
Previously, we compared all associative sections when comparing two comdat
sections. Because usually assocative sections are not mergeable sections,
we missed a lot of mergeable sections. MSVC linker doesn't seem to check
the identity of associative sections.
This patch makes LLD to ignore associative sections when doing ICF.
llvm-svn: 288483
If a section symbol is not external, that COMDAT section should never
be merge with other sections in other compilation unit. Previously,
we didn't take visibility into account.
Note that COMDAT sections with non-external visibility makes sense
because they can be removed by dead-stripping.
Fixes https://llvm.org/bugs/show_bug.cgi?id=25686
llvm-svn: 254578
There was a threading issue in the ICF code for COFF. That seems like
a venign bug in the sense that it doesn't produce an incorrect output,
but it oftentimes misses reducible sections. As a result, mergeable
sections could remain in outputs, which makes the output nondeterministic.
Basically the algorithm we are using for ICF is this: We group sections
so that identical sections will eventually be in the same group. Initially,
all sections are in one group. We split the group by relocation targets
until we get a convergence (if relocation targets are in different gruops,
the sections are different). Once a group is split, they will never be
merged.
Each section has a group ID. That variable itself is atomic, so there's
no threading issue at the level that we can use thread sanitizer.
The point is, when we split a group, we re-assign new group IDs to group
of sections. That are multiple separate writes to atomic varaibles.
Thus, splitting a group is not an atomic operation, and there's a small
chance that the other thread observes inconsistent group IDs.
Over-splitting is always "safe", so it will never create incorrect output.
I suspect that the nondeterminism stems from that point. However, I
cannot prove or fix that at this moment, so I'm going to avoid using
threads here.
llvm-svn: 251300
There's actually a room to improve this patch. Instead of not merging
sections that have different alignements, we can choose the section that
has the largest alignment requirement among all sections that are otherwise
considered the same. Then all section alignments are satisfied, so we can
merge them.
I don't know if that improvement could make any difference for real-world
input, so I'll leave it alone. Would be interesting to revisit later.
llvm-svn: 248581
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
Peter Collingbourne