Imagine that we have sections A, B, C, where A == C and
symbol ordering file containing symbols: symC, symB, symA
Previously because of ICF it was possible that final order would be
B, A or B, C. That violates order specified in ordering file.
Patch changes that.
Differential revision: https://reviews.llvm.org/D43234
llvm-svn: 326179
It should be possible to resolve undefined symbols in dynamic libraries
using symbols defined in a linker script.
Differential Revision: https://reviews.llvm.org/D43011
llvm-svn: 326176
This fixes pr36475.
I think this code can be simplified a bit, but I would like to check
in the more direct fix if we are in agreement on the direction and
then refactor.
This is not something that bfd does. The issue is not noticed in bfd
because it keeps fewer sections from the linkerscript in the output.
The reasons why it seems reasonable to do this:
- As George noticed, we would still keep the flags if the output
section had both an empty synthetic section and a regular section
- We need an heuristic to find the flags of output sections. Using the
flags of a synthetic section that would have been there seems a
reasonable heuristic.
llvm-svn: 326137
MIPS ABIs require that if an executable file uses non-PIC model, the
EI_ABIVERSION entry in the ELF header should be incremented from 0 to 1.
That allows obsoleted / limited dynamic linkers refuse to link them.
llvm-svn: 325890
This continues direction started in D43069.
We can keep sections that are explicitly assigned to segment in script.
It helps to simplify code.
Differential revision: https://reviews.llvm.org/D43571
llvm-svn: 325887
Latest patch version now.
Original commit message:
[ELF] - Do not crash with --emit-relocs and --icf=all together.
Previously we would crash because did not mark .rel[a] sections
as dead and they tried to access parent which was not live
after ICF and therefore was null.
Differential revision: https://reviews.llvm.org/D43241
llvm-svn: 325879
Previously we would crash because did not mark .rel[a] sections
as dead and they tried to access parent which was not live
after ICF and therefore was null.
Differential revision: https://reviews.llvm.org/D43241
llvm-svn: 325877
This is for fixing PR36297.
Issue itself is that if we have SECTIONS { .bar (a+b) : { *(.stub) } };
script and no section .stub, when LLD will remove .bar, but
produce output with undefined symbols a and b.
Differential revision: https://reviews.llvm.org/D43069
llvm-svn: 325875
We have an internal program that does't link without this patch. I don't
know of any open-source program that needs this, but there might be.
Since this patch improves compatibility with GNU linkers with a few lines
of code, I think it's worth to be committed.
The problem is about undefined symbols in DSOs. Some programs depend on
the GNU linkers' behavior that they pull out object files from archive
files to resolve undefined symbols in DSOs. We already allow that kind of
"reverse" dependency (from DSOs to the main executable) for regular
symbols, in particular, for "__progname" symbol (which is usually in
crt0.o), but that doesn't work if the symbol is in an archive file.
This patch is to make it work.
Differential Revision: https://reviews.llvm.org/D43658
llvm-svn: 325849
We sometimes need to iterate over input sections for a given
output section. It is not very convinent because we have to iterate
over section descriptions.
Patch introduces getInputSections helper, it simplifies things.
Differential revision: https://reviews.llvm.org/D43574
llvm-svn: 325763
This reverts commit r325679 that was committed without discussion.
Actually, in the discussion thread, most people opposed to have this
option in lld. Reverting that change doesn't mean that this is a
final decision, but that needs to be discussed first.
llvm-svn: 325714
Address @ruiu's post commit review comment about a value which is intended
to be a unsigned 32 bit integer as using uint32_t rather than unsigned.
llvm-svn: 325713
In r324043, --nopie was renamed to --no-pie to presumably fix a typo.
As it turns out, "nopie" wasn't a typo but the spelling used by
OpenBSD's binutils ld. Gold on the other hand spells the flag "no-pie".
(Vanilla binutils doesn't have a flag like this at all.)
Since they do the same thing, let's support both spellings.
llvm-svn: 325679
This patch provides migitation for CVE-2017-5715, Spectre variant two,
which affects the P5600 and P6600. It implements the LLD part of
-z hazardplt. Like the Clang part of this patch, I have opted for that
specific option name in case alternative migitation methods are required
in the future.
The mitigation strategy suggested by MIPS for these processors is to use
hazard barrier instructions. 'jalr.hb' and 'jr.hb' are hazard
barrier variants of the 'jalr' and 'jr' instructions respectively.
These instructions impede the execution of instruction stream until
architecturally defined hazards (changes to the instruction stream,
privileged registers which may affect execution) are cleared. These
instructions in MIPS' designs are not speculated past.
These instructions are defined by the MIPS32R2 ISA, so this mitigation
method is not compatible with processors which implement an earlier
revision of the MIPS ISA.
For LLD, this changes PLT stubs to use 'jalr.hb' and 'jr.hb'.
Reviewers: atanasyan, ruiu
Differential Revision: https://reviews.llvm.org/D43488
llvm-svn: 325647
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 profailing style in lld seem to be to not include such empty lines.
Clang-tidy/clang-format seem to handle this just fine.
Differential Revision: https://reviews.llvm.org/D43528
llvm-svn: 325629
Summary:
With D43396, no clients use the Path parameter anymore.
This is the lld side fix with D43400.
Depends on D43396 and D43400.
Reviewers: pcc
Subscribers: emaste, inglorion, arichardson, llvm-commits
Differential Revision: https://reviews.llvm.org/D43401
llvm-svn: 325620
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
Currently, archive file name is missing in this message. In general,
we should avoid constructing strings in an ad-hoc manner and instead
use toString() to get consistent output strings.
Differential Revision: https://reviews.llvm.org/D43420
llvm-svn: 325416
We are running lld tests with "--full-shutdown" option because we don't
want to call _exit() in lld if it is running tests. Regular shutdown
is needed for leak sanitizer.
This patch changes the way how we tell lld that it is running tests.
Now "--full-shutdown" is removed, and LLD_IN_TEST environment variable
is used instead.
This patch enables full shutdown on all ports, e.g. ELF, COFF and wasm.
Previously, we enabled it only for ELF.
Differential Revision: https://reviews.llvm.org/D43410
llvm-svn: 325413
There seems to be no reason to collect this list of symbols.
Also fix a bug where --exclude-libs would apply to all symbols that
appear in an archive's symbol table, even if the relevant archive
member was not added to the link.
Differential Revision: https://reviews.llvm.org/D43369
llvm-svn: 325380
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 accidentally dropped STB_WEAK bit from an undefined symbol
if there is a lazy object symbol with the same name. That caused a
compatibility issue with GNU gold.
llvm-svn: 325316
Even though it doesn't make sense, there seems to be multiple programs
in the wild that create PC-relative relocations in non-ALLOC sections.
I believe this is caused by the negligence of GNU linkers to not report
any errors for such relocations.
Currently, lld emits warnings against such relocations and exits.
So, you cannot link any program that contains wrong relocations until
you fix an issue in a program that generates wrong ELF files. It's often
impractical to fix a program because it's not always easy.
This patch relaxes the error checking and emit a warning instead.
Differential Revision: https://reviews.llvm.org/D43351
llvm-svn: 325307
When we are emitting a relocatable output, we should keep the original
symbol name including "@" part. Previously, we drop that part unconditionally
which resulted in dropping versions from symbols.
Differential Revision: https://reviews.llvm.org/D43307
llvm-svn: 325204
This patch addresses a minor compatibility issue with GNU linkers.
Previously, --export-dynamic-symbol is completely ignored if you
pass --export-dynamic together.
Differential Revision: https://reviews.llvm.org/D43266
llvm-svn: 325152
There are a number of different situations when symbols are requested
to be ordered in the --symbol-ordering-file that cannot be ordered for
some reason. To assist with identifying these symbols, and either
tidying up the order file, or the inputs, a number of warnings have
been added. As some users may find these warnings unhelpful, due to how
they use the symbol ordering file, a switch has also been added to
disable these warnings.
The cases where we now warn are:
* Entries in the order file that don't correspond to any symbol in the input
* Undefined symbols
* Absolute symbols
* Symbols imported from shared objects
* Symbols that are discarded, due to e.g. --gc-sections or /DISCARD/ linker script sections
* Multiple of the same entry in the order file
Reviewed by: rafael, ruiu
Differential Revision: https://reviews.llvm.org/D42475
llvm-svn: 325125
We sort inside output sections, so all the sections we see should be
InputSectionBase.
I noticed the patch adding callgraph based section ordering used this
type and changing this separately makes the merge easier.
llvm-svn: 325094
This adds an extra level of static safety to our use of placement
new to allocate Symbol types. It prevents the accidental addition
on a non-trivially-destructible member that could allocate and
leak memory.
From the spec: Storage occupied by trivially destructible objects
may be reused without calling the destructor.
Differential Revision: https://reviews.llvm.org/D43244
llvm-svn: 325025
In order to identify a compressed section, we check if a section name
starts with ".zdebug" or the section has SHF_COMPRESSED flag. We already
use the knowledge in this function. So hiding that check in
isCompressedELFSection doesn't make sense.
llvm-svn: 324951
When decompressing a compressed debug section, we drop SHF_COMPRESSED
flag but we didn't drop "z" in ".zdebug" section name. This patch does
that for consistency.
This change also fixes the issue that .zdebug_gnu_pubnames are not
dropped when we are creating a .gdb_index section.
llvm-svn: 324949
GNU gold doesn't print out ICF sections for -verbose. It only shows
them for -print-icf-sections. We printed out them for -verbose because
we didn't have -print-icf-sections. Now that we have the option, there's
no reason to print out for -verbose.
Differential Revision: https://reviews.llvm.org/D43100
llvm-svn: 324755
This is for compatiblity with GNU gold. GNU gold tries to resolve
symbols specified by --export-dynamic-symbol. So, if a symbol specified
by --export-dynamic-symbol is in an archive file, lld's result is
currently different from gold's.
Interestingly, that behavior is different for --dynamic-list.
I added a new test to ensure that.
Differential Revision: https://reviews.llvm.org/D43103
llvm-svn: 324752
This is a bit more verbose, but it has a few advantages.
The logic on what to do with special sections like .init_array is not
duplicated. Before we would need keep isKnownNonreorderableSection in
sync.
I think with this the call graph based sorting can be implemented by
"just" returning a new order from buildSectionOrder.
llvm-svn: 324744
Previously, multiple chunks of --print-icf-sections messages were interleaved
and didn't make sense. This is because forEachClass is multi-threaded.
llvm-svn: 324683
When you omit an argument, most options fall back to their defaults.
For example, --color-diagnostics is a synonym for --color-diagnostics=auto.
We don't have a way to specify the default choice for --build-id, so we
can't describe --build-id (without an argument) in that way.
This patch adds "fast" for the default build-id choice.
Differential Revision: https://reviews.llvm.org/D43032
llvm-svn: 324502
MIPS BFD linker puts _gp_disp symbol into DSO files and assigns zero
version definition index to it. This value means 'unversioned local
symbol' while _gp_disp is a section global symbol. We have to handle
this bug in the LLD because BFD linker is used for building MIPS
toolchain libraries.
Differential revision: https://reviews.llvm.org/D42486
llvm-svn: 324467
This is PR35740 which now crashes
because we remove unused synthetic sections incorrectly.
We can keep input section description and corresponding output
section live even if it must be empty and dead.
This results in a crash because SHF_LINK_ORDER handling code
tries to access first section which is nullptr in this case.
Patch fixes the issue.
Differential revision: https://reviews.llvm.org/D42681
llvm-svn: 324463
Previously --defsym=foo2=etext+2 would produce incorrect value
for foo2 because expressions did not work correctly with
reserved symbols, section offset was calculated wrong for them.
Fixes PR35744.
Differential revision: https://reviews.llvm.org/D42911
llvm-svn: 324461
Previously we ignored -plugin-opt=mcpu=<xxx>
and the only way to set CPU string was to pass
-mllvm -mcpu=<xxx>
Though clang may pass it with use of plugin options:
-plugin-opt=mcpu=x86-64
Since we are trying to be compatible in command line
with gold plugin, seems we should support it too.
Differential revision: https://reviews.llvm.org/D42956
llvm-svn: 324459
objects, it confuses codegen into generating pc-rel relocations for those
symbols, which leads to linker errors.
Differential Revision: https://reviews.llvm.org/D42977
llvm-svn: 324435
I tested that if I remove "elf" from the message, building a program
that uses libtool prints
checking whether to build shared libraries... no
but with this patch it still prints
checking whether to build shared libraries... yes
llvm-svn: 324428
In lld this was the only use of Config->Static where it meant anything
else other than "use .a instead of .so".
If a program turns out to not use any dynamic libraries, we should
produce the same result with and without -static.
llvm-svn: 324421
With fix:
Keep logic that ignores -plugin-opt=mcpu=x86-64 -plugin-opt=thinlto,
add checks for those to testcases.
Original commit message:
[ELF] - Use InitTargetOptionsFromCodeGenFlags/ParseCommandLineOptions for parsing LTO options.
gold plugin uses InitTargetOptionsFromCodeGenFlags +
ParseCommandLineOptions for parsing LTO options.
Patch do the same change for LLD.
Such change helps to avoid parsing/whitelisting LTO
plugin options again on linker side, what can help LLD
to automatically support new -plugin-opt=xxx options
passed.
Differential revision: https://reviews.llvm.org/D42733
llvm-svn: 324340
gold plugin uses InitTargetOptionsFromCodeGenFlags +
ParseCommandLineOptions for parsing LTO options.
Patch do the same change for LLD.
Such change helps to avoid parsing/whitelisting LTO
plugin options again on linker side, what can help LLD
to automatically support new -plugin-opt=xxx options
passed.
Differential revision: https://reviews.llvm.org/D42733
llvm-svn: 324322
When using Elf_Rela every tool should use the addend in the
relocation.
We have --apply-dynamic-relocs to work around bugs in tools that don't
do that.
The default value of --apply-dynamic-relocs should be false to make
sure these bugs are more easily found in the future.
llvm-svn: 324264
When resolving dynamic RELA relocations the addend is taken from the
relocation and not the place being relocated. Accordingly lld does not
write the addend field to the place like it would for a REL relocation.
Unfortunately there is some system software, in particlar dynamic loaders
such as Bionic's linker64 that use the value of the place prior to
relocation to find the offset that they have been loaded at. Both gold
and bfd control this behavior with the --[no-]apply-dynamic-relocs option.
This change implements the option and defaults it to true for compatibility
with gold and bfd.
Differential Revision: https://reviews.llvm.org/D42797
llvm-svn: 324221
We did not report valid filename for duplicate symbol error when
symbol came from binary input file.
Patch fixes it.
Differential revision: https://reviews.llvm.org/D42635
llvm-svn: 324217
Initially LLD generates Elf_Rel relocations for O32 ABI and Elf_Rela
relocations for N32 / N64 ABIs. In other words, format of input and
output relocations was always the same. Now LLD generates all output
relocations using Elf_Rel format only. It conforms to ABIs requirement.
The patch suggested by Alexander Richardson.
llvm-svn: 324064
--nopie was a typo. GNU gold doesn't recognize it. It is also
inconsistent with other options that have --foo and --no-foo.
Differential Revision: https://reviews.llvm.org/D42825
llvm-svn: 324043
In GNU linkers, the last semicolon is optional. We can't link libstdc++
with lld because of that difference.
Differential Revision: https://reviews.llvm.org/D42820
llvm-svn: 324036
Currently ICF information is output through stderr if the "--verbose"
flag is used. This differs to Gold for example, which uses an explicit
flag to output this to stdout. This commit adds the
"--print-icf-sections" and "--no-print-icf-sections" flags and changes
the output message format for clarity and consistency with
"--print-gc-sections". These messages are still output to stderr if
using the verbose flag. However to avoid intermingled message output to
console, this will not occur when the "--print-icf-sections" flag is
used.
Existing tests have been modified to expect the new message format from
stderr.
Patch by Owen Reynolds.
Differential Revision: https://reviews.llvm.org/D42375
Reviewers: ruiu, rafael
Reviewed by:
llvm-svn: 323976
Summary:
While trying to make a linker script behave the same way with lld as it did
with bfd, I discovered that lld currently doesn't diagnose overlapping
output sections. I was getting very strange runtime failures which I
tracked down to overlapping sections in the resulting binary. When linking
with ld.bfd overlapping output sections are an error unless
--noinhibit-exec is passed and I believe lld should behave the same way
here to avoid surprising crashes at runtime.
The patch also uncovered an errors in the tests: arm-thumb-interwork-thunk
was creating a binary where .got.plt was placed at an address overlapping
with .got.
Reviewers: ruiu, grimar, rafael
Reviewed By: ruiu
Differential Revision: https://reviews.llvm.org/D41046
llvm-svn: 323856
When there is a duplicate absolute symbol, LLD reports <internal>
instead of known object file name currently.
Patch fixes the issue.
Differential revision: https://reviews.llvm.org/D42636
llvm-svn: 323849
Previously an empty CPU string was passed to the LTO engine which
resulted in a generic CPU for which certain features like NOPL were
disabled. This fixes that.
Patch by Pratik Bhatu!
llvm-svn: 323801
Currently symbols assigned or created by linkerscript are not processed early
enough. As a result it is not possible to version them or assign any other flags/properties.
Patch creates Defined symbols for -defsym and linkerscript symbols early,
so that issue from above can be addressed.
It is based on Rafael Espindola's version of D38239 patch.
Fixes PR34121.
Differential revision: https://reviews.llvm.org/D41987
llvm-svn: 323729
r323476 added support for DW_FORM_line_strp, and incorrectly made that
depend on having a DWARFUnit available. We shouldn't be tracking
.debug_line_str in DWARFUnit after all. After this patch, I can do an
NFC follow up and undo a bunch of the "plumbing" part of r323476.
Differential Revision: https://reviews.llvm.org/D42609
llvm-svn: 323691
This should fix PR36017.
The root problem is that we were creating a PT_LOAD just for the
header. That was technically valid, but inconvenient: we should not be
making the ELF discontinuous.
The solution is to allow a section with LMAExpr to be added to a
PT_LOAD if that PT_LOAD doesn't already have a LMAExpr.
llvm-svn: 323625
If two sections are in the same PT_LOAD, their relatives offsets,
virtual address and physical addresses are all the same.
I initially wanted to have a single global LMAOffset, on the
assumption that every ELF file was in practiced loaded contiguously in
both physical and virtual memory.
Unfortunately that is not the case. The linux kernel has:
LOAD 0x200000 0xffffffff81000000 0x0000000001000000 0xced000 0xced000 R E 0x200000
LOAD 0x1000000 0xffffffff81e00000 0x0000000001e00000 0x15f000 0x15f000 RW 0x200000
LOAD 0x1200000 0x0000000000000000 0x0000000001f5f000 0x01b198 0x01b198 RW 0x200000
LOAD 0x137b000 0xffffffff81f7b000 0x0000000001f7b000 0x116000 0x1ec000 RWE 0x200000
The delta for all but the third PT_LOAD is the same:
0xffffffff80000000. I think the 3rd one is a hack for implementing per
cpu data, but we can't break that.
llvm-svn: 323456
This fixes the crash reported at PR36083.
The issue is that we were trying to put all the sections in the same
PT_LOAD and crashing trying to write past the end of the file.
This also adds accounting for used space in LMARegion, without it all
3 PT_LOADs would have the same physical address.
llvm-svn: 323449
Since SyntheticSection::getParent() may return null, dereferencing
this pointer in ARMExidxSentinelSection::empty() call from
removeUnusedSyntheticSections() results in crashes when linking ARM
binaries.
Patch by vit9696!
llvm-svn: 323366
The reinterpret cast to uint32_t to read the little-endian instructions
will only work on a little endian system. Use ulittle32_t to always read
little-endian (AArch64 instructions are always little endian).
Fixes PR36056
Differential Revision: https://reviews.llvm.org/D42421
llvm-svn: 323243
Summary:
First, we need to explain the core of the vulnerability. Note that this
is a very incomplete description, please see the Project Zero blog post
for details:
https://googleprojectzero.blogspot.com/2018/01/reading-privileged-memory-with-side.html
The basis for branch target injection is to direct speculative execution
of the processor to some "gadget" of executable code by poisoning the
prediction of indirect branches with the address of that gadget. The
gadget in turn contains an operation that provides a side channel for
reading data. Most commonly, this will look like a load of secret data
followed by a branch on the loaded value and then a load of some
predictable cache line. The attacker then uses timing of the processors
cache to determine which direction the branch took *in the speculative
execution*, and in turn what one bit of the loaded value was. Due to the
nature of these timing side channels and the branch predictor on Intel
processors, this allows an attacker to leak data only accessible to
a privileged domain (like the kernel) back into an unprivileged domain.
The goal is simple: avoid generating code which contains an indirect
branch that could have its prediction poisoned by an attacker. In many
cases, the compiler can simply use directed conditional branches and
a small search tree. LLVM already has support for lowering switches in
this way and the first step of this patch is to disable jump-table
lowering of switches and introduce a pass to rewrite explicit indirectbr
sequences into a switch over integers.
However, there is no fully general alternative to indirect calls. We
introduce a new construct we call a "retpoline" to implement indirect
calls in a non-speculatable way. It can be thought of loosely as
a trampoline for indirect calls which uses the RET instruction on x86.
Further, we arrange for a specific call->ret sequence which ensures the
processor predicts the return to go to a controlled, known location. The
retpoline then "smashes" the return address pushed onto the stack by the
call with the desired target of the original indirect call. The result
is a predicted return to the next instruction after a call (which can be
used to trap speculative execution within an infinite loop) and an
actual indirect branch to an arbitrary address.
On 64-bit x86 ABIs, this is especially easily done in the compiler by
using a guaranteed scratch register to pass the target into this device.
For 32-bit ABIs there isn't a guaranteed scratch register and so several
different retpoline variants are introduced to use a scratch register if
one is available in the calling convention and to otherwise use direct
stack push/pop sequences to pass the target address.
This "retpoline" mitigation is fully described in the following blog
post: https://support.google.com/faqs/answer/7625886
We also support a target feature that disables emission of the retpoline
thunk by the compiler to allow for custom thunks if users want them.
These are particularly useful in environments like kernels that
routinely do hot-patching on boot and want to hot-patch their thunk to
different code sequences. They can write this custom thunk and use
`-mretpoline-external-thunk` *in addition* to `-mretpoline`. In this
case, on x86-64 thu thunk names must be:
```
__llvm_external_retpoline_r11
```
or on 32-bit:
```
__llvm_external_retpoline_eax
__llvm_external_retpoline_ecx
__llvm_external_retpoline_edx
__llvm_external_retpoline_push
```
And the target of the retpoline is passed in the named register, or in
the case of the `push` suffix on the top of the stack via a `pushl`
instruction.
There is one other important source of indirect branches in x86 ELF
binaries: the PLT. These patches also include support for LLD to
generate PLT entries that perform a retpoline-style indirection.
The only other indirect branches remaining that we are aware of are from
precompiled runtimes (such as crt0.o and similar). The ones we have
found are not really attackable, and so we have not focused on them
here, but eventually these runtimes should also be replicated for
retpoline-ed configurations for completeness.
For kernels or other freestanding or fully static executables, the
compiler switch `-mretpoline` is sufficient to fully mitigate this
particular attack. For dynamic executables, you must compile *all*
libraries with `-mretpoline` and additionally link the dynamic
executable and all shared libraries with LLD and pass `-z retpolineplt`
(or use similar functionality from some other linker). We strongly
recommend also using `-z now` as non-lazy binding allows the
retpoline-mitigated PLT to be substantially smaller.
When manually apply similar transformations to `-mretpoline` to the
Linux kernel we observed very small performance hits to applications
running typical workloads, and relatively minor hits (approximately 2%)
even for extremely syscall-heavy applications. This is largely due to
the small number of indirect branches that occur in performance
sensitive paths of the kernel.
When using these patches on statically linked applications, especially
C++ applications, you should expect to see a much more dramatic
performance hit. For microbenchmarks that are switch, indirect-, or
virtual-call heavy we have seen overheads ranging from 10% to 50%.
However, real-world workloads exhibit substantially lower performance
impact. Notably, techniques such as PGO and ThinLTO dramatically reduce
the impact of hot indirect calls (by speculatively promoting them to
direct calls) and allow optimized search trees to be used to lower
switches. If you need to deploy these techniques in C++ applications, we
*strongly* recommend that you ensure all hot call targets are statically
linked (avoiding PLT indirection) and use both PGO and ThinLTO. Well
tuned servers using all of these techniques saw 5% - 10% overhead from
the use of retpoline.
We will add detailed documentation covering these components in
subsequent patches, but wanted to make the core functionality available
as soon as possible. Happy for more code review, but we'd really like to
get these patches landed and backported ASAP for obvious reasons. We're
planning to backport this to both 6.0 and 5.0 release streams and get
a 5.0 release with just this cherry picked ASAP for distros and vendors.
This patch is the work of a number of people over the past month: Eric, Reid,
Rui, and myself. I'm mailing it out as a single commit due to the time
sensitive nature of landing this and the need to backport it. Huge thanks to
everyone who helped out here, and everyone at Intel who helped out in
discussions about how to craft this. Also, credit goes to Paul Turner (at
Google, but not an LLVM contributor) for much of the underlying retpoline
design.
Reviewers: echristo, rnk, ruiu, craig.topper, DavidKreitzer
Subscribers: sanjoy, emaste, mcrosier, mgorny, mehdi_amini, hiraditya, llvm-commits
Differential Revision: https://reviews.llvm.org/D41723
llvm-svn: 323155
Compiler doesn't know the fact that Config->WordSize * 8 is always a
power of two, so it had to use the div instruction to divide some
number with C.
llvm-svn: 323014
I created https://reviews.llvm.org/D42202 to see how large the bloom
filter should be. With that patch, I tested various bloom filter sizes
with the following commands:
$ cmake -GNinja -DCMAKE_BUILD_TYPE=Debug -DLLVM_ENABLE_LLD=true \
-DLLVM_ENABLE_PROJECTS='clang;lld' -DBUILD_SHARED_LIBS=ON \
-DCMAKE_SHARED_LINKER_FLAGS=-Wl,-bloom-filter-bits=<some integer> \
../llvm-project/llvm
$ rm -f $(find . -name \*.so.7.0.0svn)
$ ninja lld
$ LD_BIND_NOW=1 perf stat bin/ld.lld
Here is the result:
-bloom-filter-bits=8 0.220351609 seconds
-bloom-filter-bits=10 0.217146597 seconds
-bloom-filter-bits=12 0.206870826 seconds
-bloom-filter-bits=16 0.209456312 seconds
-bloom-filter-bits=32 0.195092075 seconds
Currently we allocate 8 bits for a symbol, but according to the above
result, that number is not optimal. Even though the numbers follow the
diminishing return rule, the point where a marginal improvement becomes
too small is not -bloom-filter-bits=8 but 12. So this patch sets it to 12.
Differential Revision: https://reviews.llvm.org/D42204
llvm-svn: 323010
We need to decompose relocation type for N32 / N64 ABI. Let's do it
before any other manipulations with relocation type in the `relocateOne`
routine.
llvm-svn: 322860
The problem we had with it is that anything inside an AT is an
expression, so we failed to parse the section name because of the - in
it.
llvm-svn: 322801
Previously we always handled -defsym after other commands in command line.
That made impossible to overload values set by -defsym from linker script:
test.script:
foo = 0x22;
-defsym=foo=0x11 -script t.script
would always set foo to 0x11.
That is inconstent with common logic which allows to override command line
options. it is inconsistent with bfd behavior and seems breaks assumption that
-defsym is the same as linker script assignment, as -defsyms always handled out of
command line order.
Patch fixes the handling order.
Differential revision: https://reviews.llvm.org/D42054
llvm-svn: 322625
Symbol had both Visibility and getVisibility() and they had different
meanings. That is just too easy to get wrong.
getVisibility() would compute the visibility of a particular symbol
(foo in bar.o), and Visibility stores the computed value we will put
in the output.
There is only one case when we want what getVisibility() provides, so
inline it.
llvm-svn: 322590
We track both the combined visibility that will be used for the output
symbol and the original input visibility of the selected symbol.
Almost everything should use the computed visibility.
I will make the names less confusing an a followup patch.
llvm-svn: 322576
parseInt assumed that it could take a negative number literal (e.g.
"-123"). However, such number is in reality already handled as a
unary operator '-' followed by a number literal, so the number
literal is always non-negative. Thus, this code is dead.
llvm-svn: 322453
When a section placement (AT) command references the section itself,
the physical address of the section in the ELF header was calculated
incorrectly due to alignment happening right after the location
pointer's value was captured.
The problem was diagnosed and the first version of the patch written
by Erick Reyes.
llvm-svn: 322421
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
AT> lma_region expression allows to specify the memory region
for section load address.
Should fix PR35684.
Differential revision: https://reviews.llvm.org/D41397
llvm-svn: 322359
Summary:
As reported in bug 35788, rL316280 reintroduces a race between two
members of SectionPiece, which share the same 64 bit memory location.
To fix the race, check the hash before checking the Live member, as
suggested by Rafael.
Reviewers: ruiu, rafael
Reviewed By: ruiu
Subscribers: smeenai, emaste, llvm-commits
Differential Revision: https://reviews.llvm.org/D41884
llvm-svn: 322264
When setting up the chain, we copy over the bucket's previous symbol
index, assuming that this index will be 0 (STN_UNDEF) for an unused
bucket (marking the end of the chain). When linking with --no-rosegment,
however, unused buckets will in fact contain the padding value, and so
the hash table will end up containing invalid chains. Zero out the hash
table section explicitly to avoid this, similar to what's already done
for GNU hash sections.
Differential Revision: https://reviews.llvm.org/D41928
llvm-svn: 322259
When we have --icf=safe we should be able to define --icf=all as a
shorthand for --icf=safe --ignore-function-address-equality.
For now --ignore-function-address-equality is used only to control
access to non preemptable symbols in shared libraries.
llvm-svn: 322152
Summary:
All other templated methods have explicit instantiations but this one is
missing. Discovered while building with a clang with inliner
modifications.
Reviewers: espindola
Subscribers: emaste, llvm-commits, davidxl
Differential Revision: https://reviews.llvm.org/D41847
llvm-svn: 322057
This splits relocation processing in two steps.
First, analyze what needs to be done at the relocation spot. This can
be a constant (non preemptible symbol, relative got reference, etc) or
require a dynamic relocation. At this step we also consider creating
copy relocations.
Once that is done we decide if we need a got or a plt entry.
The code is simpler IMHO. For example:
- There is a single call to isPicRel since the logic is not split
among adjustExpr and the caller.
- R_MIPS_GOTREL is simple to handle now.
- The tracking of what is preemptible or not is much simpler now.
This also fixes a regression with symbols being both in a got and copy
relocated. They had regressed in r268668 and r268149.
The other test changes are because of error messages changes or the
order of two relocations in the output.
llvm-svn: 322047
Currently LLVM's paralellForEach has a problem with reentracy.
That caused https://bugs.llvm.org/show_bug.cgi?id=35788 (lld somtimes
hangs while linking Ruby 2.4) because maybeCompress calls writeTo which
uses paralellForEach.
This patch is to avoid using paralellForEach to call maybeCompress
to workaround the issue.
llvm-svn: 322041
The body of the in scanRelocs is fairly big. This moves it to its own
function.
It is not a big readability win by itself, but should help further
refactoring.
llvm-svn: 322035
Previously, in r320472, I moved the calculation of section offsets and sizes
for compressed debug sections into maybeCompress, which happens before
assignAddresses, so that the compression had the required information. However,
I failed to take account of relocations that patch such sections. This had two
effects:
1. A race condition existed when a debug section referred to a different debug
section (see PR35788).
2. References to symbols in non-debug sections would be patched incorrectly.
This is because the addresses of such symbols are not calculated until after
assignAddresses (this was a partial regression caused by r320472, but they
could still have been broken before, in the event that a custom layout was used
in a linker script).
assignAddresses does not need to know about the output section size of
non-allocatable sections, because they do not affect the value of Dot. This
means that there is no longer a reason not to support custom layout of
compressed debug sections, as far as I'm aware. These two points allow for
delaying when maybeCompress can be called, removing the need for the loop I
previously added to calculate the section size, and therefore the race
condition. Furthermore, by delaying, we fix the issues of relocations getting
incorrect symbol values, because they have now all been finalized.
llvm-svn: 321986
LLD previously used to handle dynamic lists and version scripts in the
exact same way, even though they have very different semantics for
shared libraries and subtly different semantics for executables. r315114
untangled their semantics for executables (building on previous work to
correct their semantics for shared libraries). With that change, dynamic
lists won't set the default version to VER_NDX_LOCAL, and so resetting
the version to VER_NDX_GLOBAL in scanShlibUndefined is unnecessary.
This was causing an issue because version scripts containing `local: *`
work by setting the default version to VER_NDX_LOCAL, but scanShlibUndefined
would override this default, and therefore symbols which should have
been local would end up in the dynamic symbol table, which differs from
both bfd and gold's behavior. gold silently keeps the symbol hidden in
such a scenario, whereas bfd issues an error. I prefer bfd's behavior
and plan to implement that in LLD in a follow-up (and the test case
added here will be updated accordingly).
Differential Revision: https://reviews.llvm.org/D41639
llvm-svn: 321982
We normally want to ignore SHT_NOBITS sections when computing
offsets. The sh_offset of section itself seems to be irrelevant and
- If the section is in the middle of a PT_LOAD, it will make no
difference on the computed offset of the followup section.
- If it is in the end of a PT_LOAD, we want to avoid its alignment
changing the offset of the followup sections.
The issue is if it is at the start of the PT_LOAD. In that case we do
have to align it so that the following sections have congruent address
and offset module the page size. We were not handling this case.
This should fix freebsd kernel link.
llvm-svn: 321657
This is "Bug 35751 - .dynamic relocation entries omitted if output
contains only IFUNC relocations"
We have InX::RelaPlt and InX::RelaIPlt synthetic sections for PLT relocations.
They are usually live in rela.plt section. Problem appears when InX::RelaPlt
section is empty. In that case we did not produce normal set of dynamic tags
required, because logic was written in the way assuming we always have
non-IRelative relocations in rela.plt.
Patch fixes the issue.
Differential revision: https://reviews.llvm.org/D41592
llvm-svn: 321600
This was raised in comments for D41592.
With current code we always assign parent
section for Rel[a] sections like
InX::RelaPlt or InX::RelaDyn, so checking
their parent for null is excessive.
llvm-svn: 321581
If using a version script with a `local: *` in it, symbols in shared
libraries will still get default visibility if another shared library on
the link line has an undefined reference to the symbol. This is quite
surprising. Neither bfd nor gold have this behavior when linking a
shared library, and none of LLD's tests fail without this behavior, so
it seems safe to limit scanShlibUndefined to executables.
As far as executables are concerned, gold doesn't do any automatic
default visibility marking, and bfd issues a link error about a shared
library having a reference to a hidden symbol rather than silently
giving that symbol default visibility. I think bfd's behavior here is
preferable to LLD's, but that's something to be considered in a
follow-up.
Differential Revision: https://reviews.llvm.org/D41524
llvm-svn: 321578
This makes adjustExpr a bit simpler too IMHO.
It seems that some of the complication around relocation processing
is that we are trying to create copy relocations too early. It seems
we could handle a few simple cases first and continue.
llvm-svn: 321507
Previously we failed to resolve them when produced executables:
"relocation R_X86_64_32 cannot be used against shared object; recompile with -fPIC"
Patch fixes it so that we resolve them to 0 for executables.
And for -shared case we still should produce the relocation.
This finishes fixing PR35720.
DIfferential revision: https://reviews.llvm.org/D41551
llvm-svn: 321473
This is an aesthetic change to represent a placeholder for later
binary patching as "0, 0, 0, 0" instead of "0x00, 0x00, 0x00, 0x00".
The former is how we represent it in COFF, and I found it easier to
read than the latter.
llvm-svn: 321471
If a relocation cannot be implemented by the dynamic linker and the
section is rw, allow creating a plt entry to use as the function
address as if the section was ro.
This matches bfd and gold. It also matches our behavior with -z
notext.
llvm-svn: 321430
Advance the memory region offset when handling a linker script data
command such as BYTE or LONG. Failure to advance the offset results
in corrupted output with overlapping sections.
Update tests to check for this combination of both a) memory regions
and b) data commands.
Fixes https://bugs.llvm.org/show_bug.cgi?id=35565
Patch by Owen Shaw!
llvm-svn: 321418
We normally avoid "switch (Config->EKind)", but in this case I think
it is worth it.
It is only executed when there is an error and it allows detemplating
a lot of code.
llvm-svn: 321404
This is part of PR35720.
Currently LLD allows dynamic relocations against text when -z notext is given.
Though for non-PIC relocations like R_X86_64_PC32 that does not work,
we produce "relocation R_X86_64_PC32 cannot be used against shared object;"
error because they may overflow in runtime.
Solution implemented is to use PLT for them.
Differential revision: https://reviews.llvm.org/D41541
llvm-svn: 321400
When two linker script symbols are subtracted, the result should be absolute.
This is the behavior of binutils' ld.
Patch by Erick Reyes!
llvm-svn: 321390
Rafael Espindola