LLD already produces a nice error message when sections exceed 4GB, and
this setRVA assertion causes LLD to crash instead of diagnosing the
error properly.
No test because we don't want slow tests that create 4GB files.
This is a different approach from D98993 that should achieve most of the
same benefit. The two changes are:
1. Sort the list of associated child sections by section name
2. Do not consider associated sections to have children themselves
This fixes the main issue, which was that we sometimes considered an
.xdata section to have a child .pdata section. That lead to slow links
and larger binaries (less xdata folding).
Otherwise, this should be NFC: we go back to ignoring .debug/.gljmp and
other metadata sections rather than only looking at pdata/xdata. We
discovered that we do care about other associated sections, like ASan
global registration metadata.
The only known reason why ICF should not merge otherwise identical
sections with differing associated sections has to do with exception
handling tables. It's not clear what ICF should do when there are other
kinds of associated sections. In every other case when this has come up,
debug info and CF guard metadata, we have opted to make ICF ignore the
associated sections.
For comparison, ELF doesn't do anything for comdat groups. Instead,
.eh_frame is parsed to figure out if a section has an LSDA, and if so,
ICF is disabled.
Another issue is that the order of associated sections is not defined.
We have had issues in the past (crbug.com/1144476) where changing the
order of the .xdata/.pdata sections in the object file lead to large ICF
slowdowns.
To address these issues, I decided it would be best to explicitly
consider only .pdata and .xdata sections during ICF. This makes it easy
to ignore the object file order, and I think it makes the intention of
the code clearer.
I've also made the children() accessor return an empty list for
associated sections. This mostly only affects ICF and GC. This was the
behavior before I made this a linked list, so the behavior change should
be good. This had positive effects on chrome.dll: more .xdata sections
were merged that previously could not be merged because they were
associated with distinct .pdata sections.
Reviewed By: mstorsjo
Differential Revision: https://reviews.llvm.org/D98993
This reverts commit bacf9cf2c5 and
reinstates commit 1a9bd5b813.
Reverting this commit did not appear to make the problem go away, so we
can go ahead and reland it.
This reverts commit 5b7aef6eb4 and relands
6529d7c5a4.
The ASan error was debugged and determined to be the fault of an invalid
object file input in our test suite, which was fixed by my last change.
LLD's project policy is that it assumes input objects are valid, so I
have added a comment about this assumption to the relocation bounds
check.
This is a pretty classic optimization. Instead of processing symbol
records and copying them to temporary storage, do a first pass to
measure how large the module symbol stream will be, and then copy the
data into place in the PDB file. This requires defering relocation until
much later, which accounts for most of the complexity in this patch.
This patch avoids copying the contents of all live .debug$S sections
into heap memory, which is worth about 20% of private memory usage when
making PDBs. However, this is not an unmitigated performance win,
because it can be faster to read dense, temporary, heap data than it is
to iterate symbol records in object file backed memory a second time.
Results on release chrome.dll:
peak mem: 5164.89MB -> 4072.19MB (-1,092.7MB, -21.2%)
wall-j1: 0m30.844s -> 0m32.094s (slightly slower)
wall-j3: 0m20.968s -> 0m20.312s (slightly faster)
wall-j8: 0m19.062s -> 0m17.672s (meaningfully faster)
I gathered similar numbers for a debug, component build of content.dll
in Chrome, and the performance impact of this change was in the noise.
The memory usage reduction was visible and similar.
Because of the new parallelism in the PDB commit phase, more cores makes
the new approach faster. I'm assuming that most C++ developer machines
these days are at least quad core, so I think this is a win.
Differential Revision: https://reviews.llvm.org/D94267
Heap profiling with ETW shows that LLD performs 4,053,721 heap
allocations over its lifetime, and ~800,000 of them come from
assocEquals. These vectors are created just to do a comparison, so fuse
the comparison into the loop and avoid the allocation.
ICF is overall a small portion of the time spent linking, and I did not
measure overall throughput improvements from this change above the noise
threshold. However, these show up in the heap profiler, and the work is
done, so we might as well land it if the code is clear enough.
Reviewed By: hans
Differential Revision: https://reviews.llvm.org/D79297
The alignment of ARM64 range extension thunks was fixed in
7c81649219, but ARM range extension thunks, and import
and delay import thunks also need aligning (like all code on ARM
platforms).
I'm adding a test for alignment of ARM64 import thunks - not
specifically adding tests for misalignment of all of them though.
Differential Revision: https://reviews.llvm.org/D77796
RangeExtensionThunkARM64 is created for out-of-range branches on Windows ARM64
because branch instructions has limited bits to encode target address.
Currently, RangeExtensionThunkARM64 is appended to its referencing COFF section
from object file at link time without any alignment requirement, so if size of
the preceding COFF section is not aligned to instruction boundary (4 bytes),
RangeExtensionThunkARM64 will emit thunk instructions at unaligned address
which is never a valid branch target on ARM64, and usually triggers invalid
instruction exception when branching to it.
This PR fixes it by requiring such thunks to align at 4 bytes.
Differential revision: https://reviews.llvm.org/D72473
This patch does the same thing as r365595 to other subdirectories,
which completes the naming style change for the entire lld directory.
With this, the naming style conversion is complete for lld.
Differential Revision: https://reviews.llvm.org/D64473
llvm-svn: 365730
Shaves another pointer off of SectionChunk, reducing the size from 96 to
88 bytes, down from 144 before I started working on this. Combined with
D62356, this reduced peak memory usage when linking chrome_child.dll
from 713MB to 675MB, or 5%.
Create NonSectionChunk to provide virtual dispatch to the rest of the
chunk types.
Reviewers: ruiu, aganea
Differential Revision: https://reviews.llvm.org/D62362
llvm-svn: 361667
Shaves another 8 bytes off of SectionChunk, the most commonly allocated
type in LLD.
These indices are only valid after we've assigned chunks to output
sections and removed empty sections, so do that in a new pass.
Reviewers: ruiu, aganea
Differential Revision: https://reviews.llvm.org/D62356
llvm-svn: 361657
This only needs to be done for MergeChunks, so just do that in a
separate pass in the Writer.
This is one small step towards eliminating the vtable in Chunk.
llvm-svn: 361573
The KeepUnique bit is used during ICF, which only operates on
SectionChunks, so only SectionChunks need it. This frees up a byte in
Chunk, which I plan to use in a follow-up change.
llvm-svn: 361549
The previous patch lost the call to PowerOf2Ceil, which causes LLD to
crash when handling common symbols with a non-power-of-2 size. I tweaked
the existing common.test to make the bsspad16 common symbol be 15 bytes
to add coverage for this case.
llvm-svn: 361426
Summary:
Valid section or chunk alignments are powers of 2 in the range [1,
8192]. These can be stored more canonically in log2 form to free up some
bits in Chunk. Combined with D61696, SectionChunk gets 8 bytes smaller.
Reviewers: ruiu, aganea
Subscribers: llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D61698
llvm-svn: 361206
Summary:
Prior to this change, every implementation of writeTo would add
OutputSectionOff to the output section buffer start before writing data.
Instead, do this math in the caller, so that it can be written once
instead of many times.
The output section offset is always equivalent to the difference between
the chunk RVA and the output section RVA, so we can replace the one
remaining usage of OutputSectionOff with that subtraction.
This doesn't change the size of SectionChunk because of alignment
requirements, but I will rearrange the fields in a follow-up change to
accomplish that.
Reviewers: ruiu, aganea
Subscribers: llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D61696
llvm-svn: 360376
SectionChunk is one of the most frequently allocated data structures in
LLD, since there are about four per function when optimizations and
debug info are enabled (.text, .pdata, .xdata, .debug$S).
A PE COFF file cannot be larger than 2GB, so there is an inherent limit
on the length of the section name and the number of relocations.
Decompose the ArrayRef and StringRef into pointer and size, and put them
back together in the accessors for section name and relocation list.
I plan to gather complete performance numbers later by padding
SectionChunk with dead data and measuring performance after all the size
optimizations are done.
llvm-svn: 359923
Summary:
Reorder the fields in both to use padding more efficiently, and add more
comments on the purpose of the fields.
Replace `std::vector<SectionChunk*> AssociativeChildren` with a
singly-linked list. This avoids the separate vector allocation to list
associative children, and shrinks the 3 pointers used for the typically
empty vector down to 1.
In the end, this reduces the sum of heap allocations used to link
browser_tests.exe with NO PDB by 13.10%, going from 2,248,728 KB to
1,954,071 KB of heap. These numbers exclude memory mapped files, which
are of course a significant factor in LLD's memory usage.
Reviewers: ruiu, mstorsjo, aganea
Subscribers: jdoerfert, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D59797
llvm-svn: 357535
Summary:
This avoids allocating O(#relocs) of intermediate data for each section
when range extension thunks aren't needed for that section. This also
removes a std::vector from SectionChunk, which further reduces its size.
Instead, this change adds the range extension thunk symbols to the
object files that contain sections that need extension thunks. By adding
them to the symbol table of the parent object, that means they now have
a symbol table index. Then we can then modify the original relocation,
after copying it to read-write memory, to use the new symbol table
index.
This makes linking browser_tests.exe with no PDB 10.46% faster, moving
it from 11.364s to 10.288s averaged over five runs.
Reviewers: mstorsjo, ruiu
Subscribers: aganea, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D59902
llvm-svn: 357200
On ARM64, this is normally necessary only after a module exceeds
128 MB in size (while the limit for thumb is 16 MB). For conditional
branches, the range limit is only 1 MB though (the same as for thumb),
and for the tbz instruction, the range is only 32 KB, which allows for
a test much smaller than the full 128 MB.
This fixes PR40467.
Differential Revision: https://reviews.llvm.org/D57575
llvm-svn: 352929
I need the comdat selection for PR40094. To keep the patch for that smaller,
I'm adding it here, and as a first application I'm using it to reject
associative comdats referring to earlier associative comdats. Depends on
D56929; together with that all associative comdats referring to other
associative comdats are now rejected.
Differential Revision: https://reviews.llvm.org/D56931
llvm-svn: 352254
to reflect the new license.
We understand that people may be surprised that we're moving the header
entirely to discuss the new license. We checked this carefully with the
Foundation's lawyer and we believe this is the correct approach.
Essentially, all code in the project is now made available by the LLVM
project under our new license, so you will see that the license headers
include that license only. Some of our contributors have contributed
code under our old license, and accordingly, we have retained a copy of
our old license notice in the top-level files in each project and
repository.
llvm-svn: 351636
Changes a few things I noticed while reading this code.
- fix a few typos in comments
- remove two `auto` uses where the type wasn't clear to me
- add comment saying that two sequential checks for `if (SparseChunks[SectionNumber] == PendingComdat)` are intentional
- name two parameters
No behavior change.
Differential Revision: https://reviews.llvm.org/D56677
llvm-svn: 351101
When these are accessed with load/store instructions on ARM64,
it becomes strictly necessary to have them properly aligned.
This fixes PR39228.
Differential Revision: https://reviews.llvm.org/D53128
llvm-svn: 344264
(patch by Benoit Rousseau)
This patch fixes a bug where the global variable initializers were sometimes not invoked in the correct order when it involved a C++ template instantiation.
Differential Revision: https://reviews.llvm.org/D52749
llvm-svn: 343847
This is a feature that MS link.exe lacks; it currently errors out on
such relocations, just like lld did before.
This allows linking clang.exe for ARM - practically, any image over
16 MB will likely run into the issue.
Differential Revision: https://reviews.llvm.org/D52156
llvm-svn: 342962
MinGW uses these kind of list terminator symbols for traversing
the constructor/destructor lists. These list terminators are
actual pointers entries in the lists, with the values 0 and
(uintptr_t)-1 (instead of just symbols pointing to the start/end
of the list).
(This mechanism exists in both the mingw-w64 crt startup code and
in libgcc; normally the mingw-w64 one is used, but a DLL build of
libgcc uses the libgcc one. Therefore it's not trivial to change
the mechanism without lots of cross-project synchronization and
potentially invalidating some combinations of old/new versions
of them.)
When mingw-w64 has been used with lld so far, the CRT startup object
files have so far provided these symbols, ending up with different,
incompatible builds of the CRT startup object files depending on
whether binutils or lld are going to be used.
In order to avoid the need of different configuration of the CRT startup
object files depending on what linker to be used, provide these symbols
in lld instead. (Mingw-w64 checks at build time whether the linker
provides these symbols or not.) This unifies this particular detail
between the two linkers.
This does disallow the use of the very latest lld with older versions
of mingw-w64 (the configure check for the list was added recently;
earlier it simply checked whether the CRT was built with gcc or clang),
and requires rebuilding the mingw-w64 CRT. But the number of users of
lld+mingw still is low enough that such a change should be tolerable,
and unifies this aspect of the toolchains, easing interoperability
between the toolchains for the future.
The actual test for this feature is added in ctors_dtors_priority.s,
but a number of other tests that checked absolute output addresses
are updated.
Differential Revision: https://reviews.llvm.org/D52053
llvm-svn: 342294
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
Normally, in order to reference exported data symbols from a different
DLL, the declarations need to have the dllimport attribute, in order to
use the __imp_<var> symbol (which contains an address to the actual
variable) instead of the variable itself directly. This isn't an issue
in the same way for functions, since any reference to the function without
the dllimport attribute will end up as a reference to a thunk which loads
the actual target function from the import address table (IAT).
GNU ld, in MinGW environments, supports automatically importing data
symbols from DLLs, even if the references didn't have the appropriate
dllimport attribute. Since the PE/COFF format doesn't support the kind
of relocations that this would require, the MinGW's CRT startup code
has an custom framework of their own for manually fixing the missing
relocations once module is loaded and the target addresses in the IAT
are known.
For this to work, the linker (originall in GNU ld) creates a list of
remaining references needing fixup, which the runtime processes on
startup before handing over control to user code.
While this feature is rather controversial, it's one of the main features
allowing unix style libraries to be used on windows without any extra
porting effort.
Some sort of automatic fixing of data imports is also necessary for the
itanium C++ ABI on windows (as clang implements it right now) for importing
vtable pointers in certain cases, see D43184 for some discussion on that.
The runtime pseudo relocation handler supports 8/16/32/64 bit addresses,
either PC relative references (like IMAGE_REL_*_REL32*) or absolute
references (IMAGE_REL_AMD64_ADDR32, IMAGE_REL_AMD64_ADDR32,
IMAGE_REL_I386_DIR32). On linking, the relocation is handled as a
relocation against the corresponding IAT slot. For the absolute references,
a normal base relocation is created, to update the embedded address
in case the image is loaded at a different address.
The list of runtime pseudo relocations contains the RVA of the
imported symbol (the IAT slot), the RVA of the location the relocation
should be applied to, and a size of the memory location. When the
relocations are fixed at runtime, the difference between the actual
IAT slot value and the IAT slot address is added to the reference,
doing the right thing for both absolute and relative references.
With this patch alone, things work fine for i386 binaries, and mostly
for x86_64 binaries, with feature parity with GNU ld. Despite this,
there are a few gotchas:
- References to data from within code works fine on both x86 architectures,
since their relocations consist of plain 32 or 64 bit absolute/relative
references. On ARM and AArch64, references to data doesn't consist of
a plain 32 or 64 bit embedded address or offset in the code. On ARMNT,
it's usually a MOVW+MOVT instruction pair represented by a
IMAGE_REL_ARM_MOV32T relocation, each instruction containing 16 bit of
the target address), on AArch64, it's usually an ADRP+ADD/LDR/STR
instruction pair with an even more complex encoding, storing a PC
relative address (with a range of +/- 4 GB). This could theoretically
be remedied by extending the runtime pseudo relocation handler with new
relocation types, to support these instruction encodings. This isn't an
issue for GCC/GNU ld since they don't support windows on ARMNT/AArch64.
- For x86_64, if references in code are encoded as 32 bit PC relative
offsets, the runtime relocation will fail if the target turns out to be
out of range for a 32 bit offset.
- Fixing up the relocations at runtime requires making sections writable
if necessary, with the VirtualProtect function. In Windows Store/UWP apps,
this function is forbidden.
These limitations are addressed by a few later patches in lld and
llvm.
Differential Revision: https://reviews.llvm.org/D50917
llvm-svn: 340726
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
Summary:
This change does three things:
- Try to find the file and line number of an undefined symbol
reference by reading codeview debug info.
- Try to find the name of the function or global variable with the
undefined symbol reference by searching the object file's symbol
table.
- Prints the information in the same style as the ELF linker.
Differential Revision: https://reviews.llvm.org/D45467
llvm-svn: 330235
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
Reid Kleckner