Compact unwind entries (CUEs) contain pointers to their respective
function symbols. However, during the link process, it's far more useful
to have pointers from the function symbol to the CUE than vice versa.
This diff adds that pointer in the form of `Defined::compactUnwind`.
In particular, when doing dead-stripping, we want to mark CUEs live when
their function symbol is live; and when doing ICF, we want to dedup
sections iff the symbols in that section have identical CUEs. In both
cases, we want to be able to locate the symbols within a given section,
as well as locate the CUEs belonging to those symbols. So this diff also
adds `InputSection::symbols`.
The ultimate goal of this refactor is to have ICF support dedup'ing
functions with unwind info, but that will be handled in subsequent
diffs. This diff focuses on simplifying `-dead_strip` --
`findFunctionsWithUnwindInfo` is no longer necessary, and
`Defined::isLive()` is now a lot simpler. Moreover, UnwindInfoSection no
longer has to check for dead CUEs -- we simply avoid adding them in the
first place.
Additionally, we now support stripping of dead LSDAs, which follows
quite naturally since `markLive()` can now reach them via the CUEs.
Reviewed By: #lld-macho, gkm
Differential Revision: https://reviews.llvm.org/D109944
Fixes PR51578 in practice.
Currently there's only enough room for a single thunk, which for real-life code
isn't enough. The error case only happens when there are many branch statements
very close to each other (0 or 1 instructions apart), with the function at the
finalization barrier small.
There's a FIXME on what to do if we hit this case, but that suggestion sounds
complicated to me (see end of PR51578 comment 5 for why).
Instead, just leave more room for thunks. Chromium's unit_tests links fine with
room for 3 thunks. Leave room for 100, which should fix this for most cases in
practice.
There's little cost for leaving lots of room: This slop value only determines
when we finalize sections, and we insert thunks for forward jumps into
unfinalized sections. So leaving room means we'll need a few more thunks, but
the thunk jump range is 128 MiB while a single thunk is just 12 bytes.
For Chromium's unit_tests:
With a slop of 3: thunk calls = 355418, thunks = 10903
With a slop of 100: thunk calls = 355426, thunks = 10904
Chances are 100 is enough for all use cases we'll hit in practice, but even
bumping it to 1000 would probably be fine.
Differential Revision: https://reviews.llvm.org/D108930
- Move a few variables closer to their uses, remove some completely
(no behavior change)
- Add some comments
- Make maxPotentialThunks include calls to stubs. It's possible that
an earlier call to a stub late in the stub table will need a thunk,
and that inserted thunk could push a stub earlier in the stub table
out of range. This is unlikely to happen, but usually there are
way fewer stub calls than non-stub calls, so if we're doing a
conservative approximation here we might as well do it correctly.
(For chromium's unit_tests target, 134421/242639 stub calls are
direct calls without this change, compared to 134408/242639 with
this change)
No real, meaningful behavior difference.
Differential Revision: https://reviews.llvm.org/D108924
- Don't subtract thunkSize from branchRange. Most places care about
the actual maximal branch range. Subtract thunkSize in the one place
that wants to leave room for a thunk.
- Set it to 0x800_0000 instead of 0xFF_FFFF
- Subtract 4 for the positive branch direction since it's a
two's complement 24bit number sign-extended mutiplied by 4,
so its range is -0x800_0000..+0x7FF_FFFC
- Make boundary checks include the boundary values
This doesn't make a huge difference in practice. It's preparation
for a "real" fix for PR51578 -- but it also lets the repro in comment 0
in that bug place one more thunk before hitting the TODO.
Differential Revision: https://reviews.llvm.org/D108897
The assert is harmless and thinks worked fine in builds with asserts enabled,
but it's still nice to fix the assert.
Differential Revision: https://reviews.llvm.org/D108853
Fixes the output segment name if both -rename_section and
-rename_segment are used and the post-section-rename segment
name is the same as the pre-segment-rename segment name to
match ld64's behavior.
The motivation is that segment$start$ can create section-less segments,
and this makes a corner case in the interaction between segment$start and
-rename_segment in the upcoming segment$start patch.
Differential Revision: https://reviews.llvm.org/D106766
ICF previously operated only within a given OutputSection. We would
merge all CFStrings first, then merge all regular code sections in a
second phase. This worked fine since CFStrings would never reference
regular `__text` sections. However, I would like to expand ICF to merge
functions that reference unwind info. Unwind info references the LSDA
section, which can in turn reference the `__text` section, so we cannot
perform ICF in phases.
In order to have ICF operate on InputSections spanning multiple
OutputSections, we need a way to distinguish InputSections that are
destined for different OutputSections, so that we don't fold across
section boundaries. We achieve this by creating OutputSections early,
and setting `InputSection::parent` to point to them. This is what
LLD-ELF does. (This change should also make it easier to implement the
`section$start$` symbols.)
This diff also folds InputSections w/o checking their flags, which I
think is the right behavior -- if they are destined for the same
OutputSection, they will have the same flags in the output (even if
their input flags differ). I.e. the `parent` pointer check subsumes the
`flags` check. In practice this has nearly no effect (ICF did not become
any more effective on chromium_framework).
I've also updated ICF.cpp's block comment to better reflect its current
status.
Reviewed By: #lld-macho, smeenai
Differential Revision: https://reviews.llvm.org/D105641
clang and gcc both seem to emit relocations in reverse order of
address. That means we can match relocations to their containing
subsections in `O(relocs + subsections)` rather than the `O(relocs *
log(subsections))` that our previous binary search implementation
required.
Unfortunately, `ld -r` can still emit unsorted relocations, so we have a
fallback code path for that (less common) case.
Numbers for linking chromium_framework on my 3.2 GHz 16-Core Intel Xeon W:
N Min Max Median Avg Stddev
x 20 4.04 4.11 4.075 4.0775 0.018027756
+ 20 3.95 4.02 3.98 3.985 0.020900768
Difference at 95.0% confidence
-0.0925 +/- 0.0124919
-2.26855% +/- 0.306361%
(Student's t, pooled s = 0.0195172)
Reviewed By: #lld-macho, thakis
Differential Revision: https://reviews.llvm.org/D105410
We have been creating many ConcatInputSections with identical values due
to .subsections_via_symbols. This diff factors out the identical values
into a Shared struct, to reduce memory consumption and make copying
cheaper.
I also changed `callSiteCount` from a uint32_t to a 31-bit field to save an
extra word.
All in all, this takes InputSection from 120 to 72 bytes (and
ConcatInputSection from 160 to 112 bytes), i.e. 30% size reduction in
ConcatInputSection.
Numbers for linking chromium_framework on my 3.2 GHz 16-Core Intel Xeon W:
N Min Max Median Avg Stddev
x 20 4.14 4.24 4.18 4.183 0.027548999
+ 20 4.04 4.11 4.075 4.0775 0.018027756
Difference at 95.0% confidence
-0.1055 +/- 0.0149005
-2.52211% +/- 0.356215%
(Student's t, pooled s = 0.0232803)
Reviewed By: #lld-macho, thakis
Differential Revision: https://reviews.llvm.org/D105305
This is a pretty big refactoring diff, so here are the motivations:
Previously, ICF ran after scanRelocations(), where we emitting
bind/rebase opcodes etc. So we had a bunch of redundant leftovers after
ICF. Having ICF run before Writer seems like a better design, and is
what LLD-ELF does, so this diff refactors it accordingly.
However, ICF had two dependencies on things occurring in Writer: 1) it
needs literals to be deduplicated beforehand and 2) it needs to know
which functions have unwind info, which was being handled by
`UnwindInfoSection::prepareRelocations()`.
In order to do literal deduplication earlier, we need to add literal
input sections to their corresponding output sections. So instead of
putting all input sections into the big `inputSections` vector, and then
filtering them by type later on, I've changed things so that literal
sections get added directly to their output sections during the 'gather'
phase. Likewise for compact unwind sections -- they get added directly
to the UnwindInfoSection now. This latter change is not strictly
necessary, but makes it easier for ICF to determine which functions have
unwind info.
Adding literal sections directly to their output sections means that we
can no longer determine `inputOrder` from iterating over
`inputSections`. Instead, we store that order explicitly on
InputSection. Bloating the size of InputSection for this purpose would
be unfortunate -- but LLD-ELF has already solved this problem: it reuses
`outSecOff` to store this order value.
One downside of this refactor is that we now make an additional pass
over the unwind info relocations to figure out which functions have
unwind info, since want to know that before `processRelocations()`. I've
made sure to run that extra loop only if ICF is enabled, so there should
be no overhead in non-optimizing runs of the linker.
The upside of all this is that the `inputSections` vector now contains
only ConcatInputSections that are destined for ConcatOutputSections, so
we can clean up a bunch of code that just existed to filter out other
elements from that vector.
I will test for the lack of redundant binds/rebases in the upcoming
cfstring deduplication diff. While binds/rebases can also happen in the
regular `.text` section, they're more common in `.data` sections, so it
seems more natural to test it that way.
This change is perf-neutral when linking chromium_framework.
Reviewed By: oontvoo
Differential Revision: https://reviews.llvm.org/D105044
ICF = Identical C(ode|OMDAT) Folding
This is the LLD ELF/COFF algorithm, adapted for MachO. So far, only `-icf all` is supported. In order to support `-icf safe`, we will need to port address-significance tables (`.addrsig` directives) to MachO, which will come in later diffs.
`check-{llvm,clang,lld}` have 0 regressions for `lld -icf all` vs. baseline ld64.
We only run ICF on `__TEXT,__text` for reasons explained in the block comment in `ConcatOutputSection.cpp`.
Here is the perf impact for linking `chromium_framekwork` on a Mac Pro (16-core Xeon W) for the non-ICF case vs. pre-ICF:
```
N Min Max Median Avg Stddev
x 20 4.27 4.44 4.34 4.349 0.043029977
+ 20 4.37 4.46 4.405 4.4115 0.025188761
Difference at 95.0% confidence
0.0625 +/- 0.0225658
1.43711% +/- 0.518873%
(Student's t, pooled s = 0.0352566)
```
Reviewed By: #lld-macho, int3
Differential Revision: https://reviews.llvm.org/D103292
`outSecFileOff` and the associated `getFileOffset()` accessors were
unnecessary.
For all the cases we care about, `outSecFileOff` is the same as
`outSecOff`. The only time they deviate is if there are zerofill
sections within a given segment. But since zerofill sections are always
at the end of a segment, the only sections where the two values deviate
are zerofill sections themselves. And we never actually query the
outSecFileOff of zerofill sections.
As for `getFileOffset()`, the only place it was being used was to
calculate the offset of the entry symbol. However, we can compute that
value by just taking the difference between the address of the entry
symbol and the address of the Mach-O header. In fact, this appears to be
what ld64 itself does. This difference is the same as the file offset as
long as there are no intervening zerofill sections, but since `__text`
is the first section in `__TEXT`, this never happens, so our previous
use of `getFileOffset()` was not wrong -- just inefficient.
Reviewed By: #lld-macho, thakis
Differential Revision: https://reviews.llvm.org/D104177
This is motivated by an upcoming diff in which the
WordLiteralInputSection ctor sets itself up based on the value of its
section flags. As such, it needs to be passed the `flags` value as part
of its ctor parameters, instead of having them assigned after the fact
in `parseSection()`. While refactoring code to make that possible, I
figured it would make sense for the other InputSections to also take
their initial values as ctor parameters.
Reviewed By: #lld-macho, thakis
Differential Revision: https://reviews.llvm.org/D103978
Not sure sectionType() carries its weight, but while we have it
we should use it consistently.
No behavior change.
Differential Revision: https://reviews.llvm.org/D104027
Our implementation draws heavily from LLD-ELF's, which in turn delegates
its string deduplication to llvm-mc's StringTableBuilder. The messiness of
this diff is largely due to the fact that we've previously assumed that
all InputSections get concatenated together to form the output. This is
no longer true with CStringInputSections, which split their contents into
StringPieces. StringPieces are much more lightweight than InputSections,
which is important as we create a lot of them. They may also overlap in
the output, which makes it possible for strings to be tail-merged. In
fact, the initial version of this diff implemented tail merging, but
I've dropped it for reasons I'll explain later.
**Alignment Issues**
Mergeable cstring literals are found under the `__TEXT,__cstring`
section. In contrast to ELF, which puts strings that need different
alignments into different sections, clang's Mach-O backend puts them all
in one section. Strings that need to be aligned have the `.p2align`
directive emitted before them, which simply translates into zero padding
in the object file.
I *think* ld64 extracts the desired per-string alignment from this data
by preserving each string's offset from the last section-aligned
address. I'm not entirely certain since it doesn't seem consistent about
doing this; but perhaps this can be chalked up to cases where ld64 has
to deduplicate strings with different offset/alignment combos -- it
seems to pick one of their alignments to preserve. This doesn't seem
correct in general; we can in fact can induce ld64 to produce a crashing
binary just by linking in an additional object file that only contains
cstrings and no code. See PR50563 for details.
Moreover, this scheme seems rather inefficient: since unaligned and
aligned strings are all put in the same section, which has a single
alignment value, it doesn't seem possible to tell whether a given string
doesn't have any alignment requirements. Preserving offset+alignments
for strings that don't need it is wasteful.
In practice, the crashes seen so far seem to stem from x86_64 SIMD
operations on cstrings. X86_64 requires SIMD accesses to be
16-byte-aligned. So for now, I'm thinking of just aligning all strings
to 16 bytes on x86_64. This is indeed wasteful, but implementation-wise
it's simpler than preserving per-string alignment+offsets. It also
avoids the aforementioned crash after deduplication of
differently-aligned strings. Finally, the overhead is not huge: using
16-byte alignment (vs no alignment) is only a 0.5% size overhead when
linking chromium_framework.
With these alignment requirements, it doesn't make sense to attempt tail
merging -- most strings will not be eligible since their overlaps aren't
likely to start at a 16-byte boundary. Tail-merging (with alignment) for
chromium_framework only improves size by 0.3%.
It's worth noting that LLD-ELF only does tail merging at `-O2`. By
default (at `-O1`), it just deduplicates w/o tail merging. @thakis has
also mentioned that they saw it regress compressed size in some cases
and therefore turned it off. `ld64` does not seem to do tail merging at
all.
**Performance Numbers**
CString deduplication reduces chromium_framework from 250MB to 242MB, or
about a 3.2% reduction.
Numbers for linking chromium_framework on my 3.2 GHz 16-Core Intel Xeon W:
N Min Max Median Avg Stddev
x 20 3.91 4.03 3.935 3.95 0.034641016
+ 20 3.99 4.14 4.015 4.0365 0.0492336
Difference at 95.0% confidence
0.0865 +/- 0.027245
2.18987% +/- 0.689746%
(Student's t, pooled s = 0.0425673)
As expected, cstring merging incurs some non-trivial overhead.
When passing `--no-literal-merge`, it seems that performance is the
same, i.e. the refactoring in this diff didn't cost us.
N Min Max Median Avg Stddev
x 20 3.91 4.03 3.935 3.95 0.034641016
+ 20 3.89 4.02 3.935 3.9435 0.043197831
No difference proven at 95.0% confidence
Reviewed By: #lld-macho, gkm
Differential Revision: https://reviews.llvm.org/D102964
Also adds support for live_support sections, no_dead_strip sections,
.no_dead_strip symbols.
Chromium Framework 345MB unstripped -> 250MB stripped
(vs 290MB unstripped -> 236M stripped with ld64).
Doing dead stripping is a bit faster than not, because so much less
data needs to be processed:
% ministat lld_*
x lld_nostrip.txt
+ lld_strip.txt
N Min Max Median Avg Stddev
x 10 3.929414 4.07692 4.0269079 4.0089678 0.044214794
+ 10 3.8129408 3.9025559 3.8670411 3.8642573 0.024779651
Difference at 95.0% confidence
-0.144711 +/- 0.0336749
-3.60967% +/- 0.839989%
(Student's t, pooled s = 0.0358398)
This interacts with many parts of the linker. I tried to add test coverage
for all added `isLive()` checks, so that some test will fail if any of them
is removed. I checked that the test expectations for the most part match
ld64's behavior (except for live-support-iterations.s, see the comment
in the test). Interacts with:
- debug info
- export tries
- import opcodes
- flags like -exported_symbol(s_list)
- -U / dynamic_lookup
- mod_init_funcs, mod_term_funcs
- weak symbol handling
- unwind info
- stubs
- map files
- -sectcreate
- undefined, dylib, common, defined (both absolute and normal) symbols
It's possible it interacts with more features I didn't think of,
of course.
I also did some manual testing:
- check-llvm check-clang check-lld work with lld with this patch
as host linker and -dead_strip enabled
- Chromium still starts
- Chromium's base_unittests still pass, including unwind tests
Implemenation-wise, this is InputSection-based, so it'll work for
object files with .subsections_via_symbols (which includes all
object files generated by clang). I first based this on the COFF
implementation, but later realized that things are more similar to ELF.
I think it'd be good to refactor MarkLive.cpp to look more like the ELF
part at some point, but I'd like to get a working state checked in first.
Mechanical parts:
- Rename canOmitFromOutput to wasCoalesced (no behavior change)
since it really is for weak coalesced symbols
- Add noDeadStrip to Defined, corresponding to N_NO_DEAD_STRIP
(`.no_dead_strip` in asm)
Fixes PR49276.
Differential Revision: https://reviews.llvm.org/D103324
The ELF format has the concept of merge sections (marked by SHF_MERGE),
which contain data that can be safely deduplicated. The Mach-O
equivalents are called literal sections (marked by S_CSTRING_LITERALS or
S_{4,8,16}BYTE_LITERALS). While the Mach-O format doesn't use the word
'merge', to avoid confusion, I've renamed our MergedOutputSection to
ConcatOutputSection. I believe it's a more descriptive name too.
This renaming sets the stage for {D102964}.
Reviewed By: #lld-macho, alexshap
Differential Revision: https://reviews.llvm.org/D102971
Vy Nguyen