Summary:
NameTypeEntry::Type is a bit-packed value of CU index+attributes (https://sourceware.org/gdb//onlinedocs/gdb/Index-Section-Format.html), which is named cu_index_and_attrs in a local variable in gdb/dwarf2read.c:dw2_symtab_iter_next
The new name CuIndexAndAttrs is more meaningful.
Reviewers: ruiu, dblaikie, espindola
Reviewed By: dblaikie
Subscribers: emaste, aprantl, arichardson, JDevlieghere, arphaman, llvm-commits
Differential Revision: https://reviews.llvm.org/D54481
llvm-svn: 346794
Previously, we uncompress all compressed sections before doing anything.
That works, and that is conceptually simple, but that could results in
a waste of CPU time and memory if uncompressed sections are then
discarded or just copied to the output buffer.
In particular, if .debug_gnu_pub{names,types} are compressed and if no
-gdb-index option is given, we wasted CPU and memory because we
uncompress them into newly allocated bufers and then memcpy the buffers
to the output buffer. That temporary buffer was redundant.
This patch changes how to uncompress sections. Now, compressed sections
are uncompressed lazily. To do that, `Data` member of `InputSectionBase`
is now hidden from outside, and `data()` accessor automatically expands
an compressed buffer if necessary.
If no one calls `data()`, then `writeTo()` directly uncompresses
compressed data into the output buffer. That eliminates the redundant
memory allocation and redundant memcpy.
This patch significantly reduces memory consumption (20 GiB max RSS to
15 Gib) for an executable whose .debug_gnu_pub{names,types} are in total
5 GiB in an uncompressed form.
Differential Revision: https://reviews.llvm.org/D52917
llvm-svn: 343979
When we write a struct to a mmap'ed buffer, we usually use
write16/32/64, but we didn't for VersionDefinitionSection, so
we needed to template that class.
llvm-svn: 343024
Previously, if you invoke lld's `main` more than once in the same process,
the second invocation could fail or produce a wrong result due to a stale
pointer values of the previous run.
Differential Revision: https://reviews.llvm.org/D52506
llvm-svn: 343009
These files used to contain classes and functions for .gdb_index,
but they are moved to SyntheticSections.{cpp,h}, so the name is now
irrelevant.
llvm-svn: 342299
With this patch, lld creates a .note.GNU_stack and adds that to an
output file if it is creating a re-linkable object file (i.e. if -r
is given). If we don't do this, and if you use GNU linkers as a final
linker, they create an executable whose stack area is executable,
which is considered pretty bad these days.
Differential Revision: https://reviews.llvm.org/D51400
llvm-svn: 340902
It turns out that postThunkContents() is only used for
sorting symbols in .symtab.
Though we can instead move the logic to SymbolTableBaseSection::finalizeContents(),
postpone calling it and then get rid of postThunkContents completely.
Differential revision: https://reviews.llvm.org/D49547
llvm-svn: 339413
This patch merges createGdbIndex function and GdbIndexSection's
constructor into a single static member function of the class.
This patch also change how we keep CU vectors. Previously, CuVector
and GdbSymbols were parallel arrays, but there's no reason to choose that
design. Now, CuVector is a member of GdbSymbol class.
A lot of members are removed from GdbIndexSection. Previously, it has
members that need to be kept in sync over several phases. I belive the new
design is less error-prone, and the new code is much easier to read
than before.
llvm-svn: 336743
.gdb_index sections can be very large. When you are compiling
multi-gibibyte executables, they can be larger than 1 GiB. The previous
implementation of .gdb_index seems to consume too much memory.
This patch reduces memory consumption by eliminating temporary objects.
In one experiment, memory consumption of GdbIndexSection class is
reduced from 962 MiB to 228 MiB when creating a .gdb_index of 1350 GiB.
Differential Revision: https://reviews.llvm.org/D49094
llvm-svn: 336672
I believe the only way to test this functionality is to create extremely
large object files and attempt to create a .gdb_index that is greater
than 4 GiB. But I think that's too much for most environments and buildbots,
so I'm commiting this without a test that actually triggers the new
error condition.
llvm-svn: 336631
Patch by Rahul Chaudhry!
This change adds experimental support for SHT_RELR sections, proposed
here: https://groups.google.com/forum/#!topic/generic-abi/bX460iggiKg
Pass '--pack-dyn-relocs=relr' to enable generation of SHT_RELR section
and DT_RELR, DT_RELRSZ, and DT_RELRENT dynamic tags.
Definitions for the new ELF section type and dynamic array tags, as well
as the encoding used in the new section are all under discussion and are
subject to change. Use with caution!
Pass '--use-android-relr-tags' with '--pack-dyn-relocs=relr' to use
SHT_ANDROID_RELR section type instead of SHT_RELR, as well as
DT_ANDROID_RELR* dynamic tags instead of DT_RELR*. The generated
section contents are identical.
'--pack-dyn-relocs=android+relr --use-android-relr-tags' enables both
'--pack-dyn-relocs=android' and '--pack-dyn-relocs=relr': lld will
encode the relative relocations in a SHT_ANDROID_RELR section, and pack
the rest of the dynamic relocations in a SHT_ANDROID_REL(A) section.
Differential Revision: https://reviews.llvm.org/D48247
llvm-svn: 336594
Almost all entries inside MIPS GOT are referenced by signed 16-bit
index. Zero entry lies approximately in the middle of the GOT. So the
total number of GOT entries cannot exceed ~16384 for 32-bit architecture
and ~8192 for 64-bit architecture. This limitation makes impossible to
link rather large application like for example LLVM+Clang. There are two
workaround for this problem. The first one is using the -mxgot
compiler's flag. It enables using a 32-bit index to access GOT entries.
But each access requires two assembly instructions two load GOT entry
index to a register. Another workaround is multi-GOT. This patch
implements it.
Here is a brief description of multi-GOT for detailed one see the
following link https://dmz-portal.mips.com/wiki/MIPS_Multi_GOT.
If the sum of local, global and tls entries is less than 64K only single
got is enough. Otherwise, multi-got is created. Series of primary and
multiple secondary GOTs have the following layout:
```
- Primary GOT
Header
Local entries
Global entries
Relocation only entries
TLS entries
- Secondary GOT
Local entries
Global entries
TLS entries
...
```
All GOT entries required by relocations from a single input file
entirely belong to either primary or one of secondary GOTs. To reference
GOT entries each GOT has its own _gp value points to the "middle" of the
GOT. In the code this value loaded to the register which is used for GOT
access.
MIPS 32 function's prologue:
```
lui v0,0x0
0: R_MIPS_HI16 _gp_disp
addiu v0,v0,0
4: R_MIPS_LO16 _gp_disp
```
MIPS 64 function's prologue:
```
lui at,0x0
14: R_MIPS_GPREL16 main
```
Dynamic linker does not know anything about secondary GOTs and cannot
use a regular MIPS mechanism for GOT entries initialization. So we have
to use an approach accepted by other architectures and create dynamic
relocations R_MIPS_REL32 to initialize global entries (and local in case
of PIC code) in secondary GOTs. But ironically MIPS dynamic linker
requires GOT entries and correspondingly ordered dynamic symbol table
entries to deal with dynamic relocations. To handle this problem
relocation-only section in the primary GOT contains entries for all
symbols referenced in global parts of secondary GOTs. Although the sum
of local and normal global entries of the primary got should be less
than 64K, the size of the primary got (including relocation-only entries
can be greater than 64K, because parts of the primary got that overflow
the 64K limit are used only by the dynamic linker at dynamic link-time
and not by 16-bit gp-relative addressing at run-time.
The patch affects common LLD code in the following places:
- Added new hidden -mips-got-size flag. This flag required to set low
maximum size of a single GOT to be able to test the implementation using
small test cases.
- Added InputFile argument to the getRelocTargetVA function. The same
symbol referenced by GOT relocation from different input file might be
allocated in different GOT. So result of relocation depends on the file.
- Added new ctor to the DynamicReloc class. This constructor records
settings of dynamic relocation which used to adjust address of 64kb page
lies inside a specific output section.
With the patch LLD is able to link all LLVM+Clang+LLD applications and
libraries for MIPS 32/64 targets.
Differential revision: https://reviews.llvm.org/D31528
llvm-svn: 334390
These maps are small, but we are creating an destroying one for each
input .eh_frame.
This patch reduces the total memory allocation from 765.54MB to
749.19MB. The peak is still the same: 563.7MB.
llvm-svn: 331075
Now that getSectionPiece is fast (uses a hash) it is probably OK to
split merge sections early.
The reason I want to do this is to split eh_frame sections in the same
place.
This does mean that we have to decompress early. Given that the only
compressed sections are debug info, I don't think we are missing much.
It is a small improvement: 0.5% on the geometric mean.
llvm-svn: 331058
This is slightly simpler to read IMHO. Now if a symbol has a position
in the file, it is Defined.
The main motivation is that with this a SharedSymbol doesn't need a
section, which reduces the size of SymbolUnion.
With this the peak allocation when linking chromium goes from 568.1 to
564.2 MB.
llvm-svn: 330966
MIPS ABI requires creation of the MIPS_RLD_MAP dynamic tag for non-PIE
executables only and MIPS_RLD_MAP_REL tag for both PIE and non-PIE
executables. The patch skips definition of the MIPS_RLD_MAP for PIE
files and defines MIPS_RLD_MAP_REL.
The MIPS_RLD_MAP_REL tag stores the offset to the .rld_map section
relative to the address of the tag itself.
Differential Revision: https://reviews.llvm.org/D43347
llvm-svn: 329996
Also make certain Thunk methods non-const as this will be required for
an upcoming change.
Differential Revision: https://reviews.llvm.org/D44961
llvm-svn: 328732
Choosing a Shift2 value based on wordsize is cargo-culted from gold.
Assuming that djb hash is a good hash function, choosing bits [4,9]
shouldn't be any worse or better than choosing bits [5,10]. We shouldn't
have copied that behavior that we can't justify in the first place.
Differential Revision: https://reviews.llvm.org/D44547
llvm-svn: 327921
This is the same as 327248 except Arm defining _GLOBAL_OFFSET_TABLE_ to
be the base of the .got section as some existing code is relying upon it.
For most Targets the _GLOBAL_OFFSET_TABLE_ symbol is expected to be at
the start of the .got.plt section so that _GLOBAL_OFFSET_TABLE_[0] =
reserved value that is by convention the address of the dynamic section.
Previously we had defined _GLOBAL_OFFSET_TABLE_ as either the start or end
of the .got section with the intention that the .got.plt section would
follow the .got. However this does not always hold with the current
default section ordering so _GLOBAL_OFFSET_TABLE_[0] may not be consistent
with the reserved first entry of the .got.plt.
X86, X86_64 and AArch64 will use the .got.plt. Arm, Mips and Power use .got
Fixes PR36555
Differential Revision: https://reviews.llvm.org/D44259
llvm-svn: 327823
This change broke ARM code that expects to be able to add
_GLOBAL_OFFSET_TABLE_ to the result of an R_ARM_REL32.
I will provide a reproducer on llvm-commits.
llvm-svn: 327688
the start of the .got.plt section so that _GLOBAL_OFFSET_TABLE_[0] =
reserved value that is by convention the address of the dynamic section.
Previously we had defined _GLOBAL_OFFSET_TABLE_ as either the start or end
of the .got section with the intention that the .got.plt section would
follow the .got. However this does not always hold with the current
default section ordering so _GLOBAL_OFFSET_TABLE_[0] may not be consistent
with the reserved first entry of the .got.plt.
X86, X86_64, Arm and AArch64 will use the .got.plt. Mips and Power use .got
Fixes PR36555
Differential Revision: https://reviews.llvm.org/D44259
llvm-svn: 327248
Summary:
Before the name of the function sounded like it was just a getter for the
private class member Addend. However, it actually calculates the final
value for the r_addend field in Elf_Rela that is used when writing the
.rela.dyn section. I also added a comment to the UseSymVA member to
explain how it interacts with computeAddend().
Differential Revision: https://reviews.llvm.org/D43161
llvm-svn: 325485
Now that we have R_ADDEND, UseSymVA was redundant. We only want to
write the symbol virtual address when using an expression other than
R_ADDEND.
llvm-svn: 325360
Summary:
This follows up on r321889 where writing of Elf_Rel addends was partially
moved to RelocationBaseSection. This patch ensures that the addends are
always written to the output section when a input section uses RELA but the
output is REL.
Differential Revision: https://reviews.llvm.org/D42843
llvm-svn: 325328
Previously we checked (HeaderSize == 0) to find out if
PltSection section is IPLT or PLT. Some targets does not set
HeaderSize though. For example PPC64 has no lazy binding implemented
and does not set PltHeaderSize constant.
Because of that using of both IPLT and PLT relocations worked
incorrectly there (testcase is provided).
Patch fixes the issue.
Differential revision: https://reviews.llvm.org/D41613
llvm-svn: 322362
We add dynamic section entries both in the ctor of the class and
DynamicSection::finalizeContents(). Some entries need to be added early
in the ctor because they add strings to .dynstr. Other entries were
intended to be added in finalizeContents(). However, some entries are
added in the ctor even though they don't add strings. This patch
fix the issue.
llvm-svn: 320851
This change actually makes the linker slightly faster. My observation
is that, with this patch, link time of clang without debug is about 1%
faster.
Differential Revision: https://reviews.llvm.org/D40697
llvm-svn: 319600
Common symbols are now represented with a DefinedRegular that points
to a BssSection, even during symbol resolution.
Differential Revision: https://reviews.llvm.org/D39666
llvm-svn: 317447
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
The Android relocation packing format is a more compact
format for dynamic relocations in executables and DSOs
that is based on delta encoding and SLEBs. An overview
of the format can be found in the Android source code:
https://android.googlesource.com/platform/bionic/+/refs/heads/master/tools/relocation_packer/src/delta_encoder.h
This patch implements relocation packing using that format.
This implementation uses a more intelligent algorithm for compressing
relative relocations than Android's own relocation packer. As a
result it can generally create smaller relocation sections than
that packer. If I link Chromium for Android targeting ARM32 I get a
.rel.dyn of size 174693 bytes, as compared to 371832 bytes with gold
and the Android packer.
Differential Revision: https://reviews.llvm.org/D39152
llvm-svn: 316775
Previously, EhFrameSection pushes data to EhFrameHdr by calling addFde
function. By making EhFrameHdr pull data from EhFrameSection, we can
eliminate a member from EhFrameHdr.
llvm-svn: 316730
String merging is one of the most time-consuming functions in lld.
This patch parallelize it to speed it up. On my 2-socket 20-core
40-threads Xeon E5-2680 @ 2.8 GHz machine, this patch shorten the
clang debug build link time from 7.11s to 5.16s. It's a 27%
improvement and actually pretty noticeable. In this test condition,
lld is now 4x faster than gold.
Differential Revision: https://reviews.llvm.org/D38266
llvm-svn: 314588
Convert all common symbols to regular symbols after scan.
This means that the downstream code does not to handle common symbols as a special case.
Differential Revision: https://reviews.llvm.org/D38137
llvm-svn: 314495
This is "Bug 34688 - lld much slower than bfd when linking the linux kernel"
Inside copyRelocations() we have O(N*M) algorithm, where N - amount of
relocations and M - amount of symbols in symbol table. It isincredibly slow
for linking linux kernel.
Patch creates local search tables to speedup.
With this fix link time goes for me from 12.95s to 0.55s what is almost 23x
faster. (used release LLD).
Differential revision: https://reviews.llvm.org/D38129
llvm-svn: 314282
We used to use std::set to uniquify CU vector elements, but as we know,
std::set is pretty slow. Fortunately we didn't actually have to use a
std::set here. This patch replaces it with std::vector.
With this patch, lld's -gdb-index overhead when linking a clang debug
build is now about 1 second (8.65 seconds without -gdb-index vs 9.60
seconds with -gdb-index). Since gold takes more than 6 seconds to create
a .gdb_index for the same output, our number isn't that bad.
llvm-svn: 314094
Previously, we had two levels of hash table lookup. The first hash
lookup uses CachedHashStringRefs as keys and returns offsets in string
table. Then, we did the second hash table lookup to obtain GdbSymbol
pointers. But we can directly map strings to GDbSymbols.
One test file is updated in this patch because we no longer have a '\0'
byte at the start of the string pool, which was automatically inserted
by StringTableBuilder.
This patch speeds up Clang debug build (with -gdb-index) link time by
0.3 seconds.
llvm-svn: 314092
This change alone speeds up linking of Clang debug build with -gdb-index
by 1.2 seconds, from 12.5 seconds to 11.3 seconds. (Without -gdb-index,
lld takes 8.5 seconds to link the same input files.)
llvm-svn: 314090
This patch rewrites a part of GdbIndexSection to address the following
issues in the previous implementation:
- Previously, some struct declarations were in GdbIndex.h while they
were not used in GdbIndex.cpp. Such structs are moved to
SyntheticSection.h.
- The actual implementation were split into GdbIndexSection and GdbHash
section, but that separation didn't make much sense. They are now
unified as GdbIndexSection.
In addition to the above changes, this patch splits functions, rename
variables and remove redundant functions/variables to generally improve
code quality.
llvm-svn: 314084
Its a PR34648 which was a segfault that happened because
we stored pointers to elements in DenseMap.
When DenseMap grows such pointers are invalidated.
Solution implemented is to keep elements by pointer
and not by value.
Differential revision: https://reviews.llvm.org/D38034
llvm-svn: 313741
EhSectionPiece used to have a pointer to a section, but that pointer was
mostly redundant because we almost always know what the section is without
using that pointer. This patch removes the pointer from the struct.
This patch also use uint32_t/int32_t instead of size_t to represent
offsets that are hardly be larger than 4 GiB. At the moment, I think it is
OK even if we cannot handle .eh_frame sections larger than 4 GiB.
Differential Revision: https://reviews.llvm.org/D38012
llvm-svn: 313697
CieRecord is a struct containing a CIE and FDEs, but oftentimes the
struct itself is named `Cie` which caused some confusion. This patch
renames them `CieRecords` or `Rec`.
llvm-svn: 313681
to separate commons based on file name patterns. The following linker script
construct does not work because commons are allocated before section placement
is done and the only synthesized BssSection that holds all commons has no file
associated with it:
SECTIONS { .common_0 : { *file0.o(COMMON) }}
This patch changes the allocation of commons to create a section per common
symbol and let the section logic do the layout.
Differential revision: https://reviews.llvm.org/D37489
llvm-svn: 312796
GdbIndexSection doesn't need lazy finalization because when an instance
of the class is created, we already know all debug info sections.
We can initialize the instnace in the ctor.
llvm-svn: 310931
I have a patch to let DwarfContext defer to lld for getting section
contents and relocations.
That is a pretty big performance improvement.
This is just a refactoring to make that easier to do.
This change makes the *creation* of gdb index a dedicated step and
makes that templated. That is so that we can uses Elf_Rel in the code.
llvm-svn: 307867
This patch reimplements .gdb_index in more natural way,
that makes proccess of switching to multithreaded index building to
be trivial.
Differential revision: https://reviews.llvm.org/D33552
llvm-svn: 304927
Before InputSectionBase had an OutputSection pointer, but that was not
always valid. For example, if it was a merge section one actually had
to look at MergeSec->OutSec.
This was brittle and caused bugs like the one fixed by r304260.
We now have a single Parent pointer that points to an OutputSection
for InputSection, but to a SyntheticSection for merge sections and
.eh_frame. This makes it impossible to accidentally access an invalid
OutSec.
llvm-svn: 304338
I found this when builded llc binary using gcc 5.4.1 + LLD.
gcc produces duplicate entries in .debug_gnu_pubtypes section, ex:
UnifyFunctionExitNodes.cpp.o has:
0x0000ac07 EXTERNAL TYPE "std::success_type<void*>"
0x0000ac07 EXTERNAL TYPE "std::success_type<void*>"
clang produces single entry here:
0x0000d291 EXTERNAL TYPE "std::__success_type<void *>"
If we link output from gcc with LLD, that would produce excessive duplicate
entries in .gdb_index constant pool area. That does not seem affect gdb work,
but makes .gdb_index larger than it can be.
I also checked that gold filters out such duplicates too. Patch fixes it.
Differential revision: https://reviews.llvm.org/D32647
llvm-svn: 303975
https://sourceware.org/gdb/onlinedocs/gdb/Index-Section-Format.html says:
"A CU vector in the constant pool is a sequence of offset_type values.
The first value is the number of CU indices in the vector.
Each subsequent value is the index and symbol attributes of a CU in the CU list."
Previously we keeped 2 values until the end, what was useless.
Initially was a part of D32647, though it is possible to split out.
Patch do that.
Differential revision: https://reviews.llvm.org/D33551
llvm-svn: 303973
It is not clear why a synthetic section wants to use padding defined
in the linker script. The padding is for the space between sections.
It was also missing a test.
llvm-svn: 303812
GetSection is a template because write calls relocate.
relocate has two parts. The non alloc code really has to be a
template, as it is looking a raw input file data.
The alloc part is only a template because of getSize.
This patch folds the value of getSize early, detemplates
getRelocTargetVA and splits relocate into a templated non alloc case
and a regular function for the alloc case. This has the nice advantage
of making sure we collect all the information we need for relocations
before getting to InputSection::relocateNonAlloc.
Since we know got is alloc, it can just call the function directly and
avoid the template.
llvm-svn: 303355
SymbolTableBaseSection was introduced.
Detemplation of SymbolTableSection should allow to detemplate more things.
Differential revision: https://reviews.llvm.org/D33124
llvm-svn: 303150
We generally want to use uint64_t instead of uintX_t if the 64-bit
type works for both 32-bit and 64-bit because it is simpler than
the variable-size type.
llvm-svn: 300293
This patch changes GnuHashTableSection implementation
to avoid depend on uintX_t and other ELFT stuff, reducing amount of
changes for following patch(es).
Differential revision: https://reviews.llvm.org/D31184
llvm-svn: 298998
Fangrui Song