Change some decodetype functions to operate on bytes nstead of
Symbol. This is in preparation of implementing live method
tracking in index-based deadcode pass, and reducing/eliminating
sym.Symbol in general.
Change-Id: Ia9809ad7b182884225e1bda577e8dbec0cd216c5
Reviewed-on: https://go-review.googlesource.com/c/go/+/199077
Reviewed-by: Austin Clements <austin@google.com>
As we move the debug_line generation into the compiler, we need to
upgrade the notion of compilationUnit to not just be on a per package
basis. That won't be the case as it will be impossible for all
compilationUnits to have the same set of files names used to build the
debug_lines table. (For example, assembled files in a package don't know
about any files but themselves, so the debug_lines table could only
reference themseves. As such, we need to break the 1:1 relationship
between compUnit and package.)
Change-Id: I2e517bb6c01de0115bbf777af828a2fe59c09ce8
Reviewed-on: https://go-review.googlesource.com/c/go/+/189618
Run-TryBot: Jeremy Faller <jeremy@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Austin Clements <austin@google.com>
Reviewed-by: Cherry Zhang <cherryyz@google.com>
For linking large binaries, the slice of Relocs consumes a large
amount of memory. We can reduce this memory consumption by
shrinking the size of the Reloc struct. This CL moves the fields
used only in external linking or only on PPC64 and S390X to a
lazily initialized side struct.
Linking k8s.io/kubernetes/cmd/kube-apiserver on Linux/AMD64,
before:
inuse_space 1240.25MB total
438.11MB 35.32% 35.32% 438.11MB 35.32% cmd/link/internal/objfile.(*objReader).readSlices
after:
inuse_space 1123.39MB total
306.85MB 27.31% 55.03% 306.85MB 27.31% cmd/link/internal/objfile.(*objReader).readSlices
Under GOGC=5 (to simulate system under memory pressure), the max
RSS reduces from ~2.05G to ~1.83G. Even with external linking the
max RSS doesn't increase.
Change-Id: I9bb8a8effaed24a8b3c4b1cfb93e6992b8740f91
Reviewed-on: https://go-review.googlesource.com/c/go/+/173437
Run-TryBot: Cherry Zhang <cherryyz@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
It is useful to be able to dig the Go version out of a binary,
even a stripped binary. rsc.io/goversion does this for x86
binaries by disassembling the binary to find runtime/proc.go's
startup-time reference to runtime.buildVersion's address.
That approach is quite fragile: the implementation doesn't
work for non-x86 and must be updated as the generated
code changes.
rsc.io/goversion finds the module version string by looking
for random 16-byte framing around the actual string.
This is less fragile but fairly kludgy and requires scanning
the entire data segment.
cmd/buildid finds the build ID by looking for an ELF note
or else falling back to scanning the beginning of the text
segment for a magic string. This has proved quite reliable
and doesn't require scanning much of the binary.
This CL makes it possible to find the Go and module versions
using a scan more like the build ID scan: a symbol early in
the writable data segment starts with a magic number and
then has pointers to the two string variables.
Setting up for "go version <binary>".
For #31624.
Change-Id: I78ea8c52fe1686b5cc5a829ca5f198104d10ebf0
Reviewed-on: https://go-review.googlesource.com/c/go/+/173342
Run-TryBot: Russ Cox <rsc@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Jay Conrod <jayconrod@google.com>
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
When the linker's -X flag is used, it will overwrite the symbol's
content (sym.P) in addstrdata. The symbol's content may be in
read-only memory, in which case overwriting it will fault. Do
copy-on-write to fix this.
Change-Id: I34d583f44c30d187042757e19a14c1ef7d3e613c
Reviewed-on: https://go-review.googlesource.com/c/go/+/173937
Run-TryBot: Cherry Zhang <cherryyz@google.com>
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
Reviewed-by: Than McIntosh <thanm@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
This resurrects CL 121198, except that this time we map read-only.
In case that we need to apply relocations to the symbol's
content that is backed by read-only memory, we do our own copy-
on-write. This can happen if we failed to mmap the output file,
or we build for Wasm.
Memory profile for building k8s.io/kubernetes/cmd/kube-apiserver
on Linux/AMD64:
Old (before this sequence of CLs):
inuse_space 1598.75MB total
669.87MB 41.90% 41.90% 669.87MB 41.90% cmd/link/internal/objfile.(*objReader).readSlices
New:
inuse_space 1280.45MB total
441.18MB 34.46% 34.46% 441.18MB 34.46% cmd/link/internal/objfile.(*objReader).readSlices
Change-Id: I6b4d29d6eee9828089ea3120eb38c212db21330b
Reviewed-on: https://go-review.googlesource.com/c/go/+/170741
Run-TryBot: Cherry Zhang <cherryyz@google.com>
Reviewed-by: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Move the phase of applying relocations later, after the sections
and segments are written to the mmap'd output region. Then apply
relocations directly in the output region, instead of the input.
So the input slices we read in don't need to be modified.
This is in preparation for mmap'ing input files read-only.
Change-Id: If9c80657b4469da36aec5a9ab6acf664f5af8fa0
Reviewed-on: https://go-review.googlesource.com/c/go/+/170739
Run-TryBot: Cherry Zhang <cherryyz@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Austin Clements <austin@google.com>
We are preparing for applying relocations to the output buffer.
However, for DWARF compression, relocations need to be applied
before compression, but we don't have an output buffer at that
time. We also cannot delay DWARF compression to when we mmap the
output file, because we need the size of the DWARF sections to
compute the file size.
Instead of applying all the relocations together, we apply
relocations in DWARF sections one symbol at a time, right before
it is writing out for compression. As the symbol content may be
in read-only memory (in the future), we use a temporary buffer
for applying the relocations, and immediately write it out.
If compression is not used, relocations are still applied all
together.
This is in preparation for mmap'ing input files read-only.
Change-Id: Iae6d2dd71313897d5054bcc458d3bb78075b30c3
Reviewed-on: https://go-review.googlesource.com/c/go/+/171397
Run-TryBot: Cherry Zhang <cherryyz@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Austin Clements <austin@google.com>
Apply R_DWARFFILEREF relocations later, along with other
relocations, so that we don't modify symbols' contents before
they are written to the output buffer.
This is in preparation for mmap'ing input files read-only.
Change-Id: I8e9ffb2f05acf8f198589b8770f277beb3847541
Reviewed-on: https://go-review.googlesource.com/c/go/+/170740
Run-TryBot: Cherry Zhang <cherryyz@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Austin Clements <austin@google.com>
Use mmap for writing most of the output file content,
specifically, the sections and segments. After layout, we
already know the sizes and file offsets for the sections and
segments. So we can just write the bytes by copying to a mmap'd
backing store.
The writing of the output file is split into two parts. The first
part writes the sections and segments to the mmap'd region. The
second part writes some extra content, for which we don't know
the size, so we use direct file IO.
This is in preparation for mmap'ing input files read-only.
Change-Id: I9f3b4616a9f96bfd5c940d74c50aacd6d330f7d2
Reviewed-on: https://go-review.googlesource.com/c/go/+/170738
Run-TryBot: Cherry Zhang <cherryyz@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Austin Clements <austin@google.com>
Currently the offset values of ELF relocations and Macho relocations
are 256 and 512 respectively, which means that the space reserved for
ELF relocations is only 256. But AARCH64 has more than 256 ELF relocation
types, in fact the maximum AARCH64 ELF relocation type recorded in file
src/debug/elf/elf.go is 1032 currently. So this CL increases the offset
of Macho relocations to 2048 to leave enough space for AARCH64 ELF
relocations.
Change-Id: I784ac38aeb3e102ac7825f6d621086849c8d3146
Reviewed-on: https://go-review.googlesource.com/c/go/+/172497
Run-TryBot: Brad Fitzpatrick <bradfitz@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
We're going to need a different TLS offset for Android Q, so the static
offsets used for 386 and amd64 are no longer viable on Android.
Introduce runtime·tls_g and use that for indexing into TLS storage. As
an added benefit, we can then merge the TLS setup code for all android
GOARCHs.
While we're at it, remove a bunch of android special cases no longer
needed.
Updates #29674
Updates #29249 (perhaps fixes it)
Change-Id: I77c7385aec7de8f1f6a4da7c9c79999157e39572
Reviewed-on: https://go-review.googlesource.com/c/go/+/169817
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Cherry Zhang <cherryyz@google.com>
Change-Id: Ia268b0d64dc89866aa09bfffcaa109741088a904
Reviewed-on: https://go-review.googlesource.com/c/go/+/169119
Run-TryBot: Ian Lance Taylor <iant@golang.org>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
This change optimizes the data section in the wasm binary by
omitting blocks of zeroes and instead emitting data segments
with offsets skipping the zeroes.
This optimization is inspired by the memory-packing pass of the
wasm-opt tool and reduces the wasm binary size of "hello world" by 14%.
Change-Id: Iba3043df05bf6aab4745c5f8015c0337fc218aff
Reviewed-on: https://go-review.googlesource.com/c/go/+/167801
Run-TryBot: Brad Fitzpatrick <bradfitz@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Cherry Zhang <cherryyz@google.com>
Instead of writing an init function per package that does the same
thing for every package, just write that implementation once in the
runtime. Change the compiler to generate a data structure that encodes
the required initialization operations.
Reduces cmd/go binary size by 0.3%+. Most of the init code is gone,
including all the corresponding stack map info. The .inittask
structures that replace them are quite a bit smaller.
Most usefully to me, there is no longer an init function in every -S output.
(There is an .inittask global there, but it's much less distracting.)
After this CL we could change the name of the "init.ializers" function
back to just "init".
Update #6853
R=go1.13
Change-Id: Iec82b205cc52fe3ade4d36406933c97dbc9c01b1
Reviewed-on: https://go-review.googlesource.com/c/go/+/161337
Run-TryBot: Keith Randall <khr@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Josh Bleecher Snyder <josharian@gmail.com>
This commit fixes trampoline generation on aix/ppc64 which must use TOC
symbols.
It also adds a size to runtime.text.X symbols to prevent ld from moving
them, like runtime.text.
Change-Id: Ida033ec20ad8d7b7fb3faeb0ec4fa7bc4ce86b7e
Reviewed-on: https://go-review.googlesource.com/c/go/+/164009
Run-TryBot: Ian Lance Taylor <iant@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
In order to allow DWARF with ld, the symbol table is adapted.
In internal linkmode, each package is considered as a .FILE. However,
current version of ld is crashing on a few programs because of
relocations between DWARF symbols. Considering all packages as part of
one .FILE seems to bypass this bug.
As it might be fixed in a future release, the size of each package
in DWARF sections is still retrieved and can be used when it's fixed.
Moreover, it's improving internal linkmode which should have done it
anyway.
Change-Id: If3d023fe118b24b9f0f46d201a4849eee8d5e333
Reviewed-on: https://go-review.googlesource.com/c/go/+/164006
Run-TryBot: Ian Lance Taylor <iant@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
Moduledata symbols like runtime.data or runtime.text must have the
same position in the final executable (as some symbol accesses are made
by offset from them).
ld on AIX might move them randomly if there are nil size symbols.
ld will also remove unreachable symbols like runtime.epclntab or
runtime.rodata. In order to keep them, R_REF relocations are created
between firstmoduledata and these symbols. This relocation tells ld to
keep these symbols even if there aren't reachable.
Change-Id: Ie5a28cf406977131cec6442f7f5b6fd89fb775a3
Reviewed-on: https://go-review.googlesource.com/c/go/+/164004
Run-TryBot: Ian Lance Taylor <iant@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
This commit adds external linking in cmd/link for aix/ppc64.
As relocations on .text data aren't possible on AIX, Segrelrodata is
used to move all these datas to .data section.
Change-Id: I4d1361c1fc9290e11e6f5560864460c76551dbeb
Reviewed-on: https://go-review.googlesource.com/c/go/+/164003
Run-TryBot: Tobias Klauser <tobias.klauser@gmail.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
This commit fixes the size of outer symbols like type.*.
Outer symbols cannot have a nil size on AIX or they will be
removed by ld as long as all their sub-symbols.
Change-Id: I68ff3ce5a3a034e3c3eb23431aba31245073cf20
Reviewed-on: https://go-review.googlesource.com/c/163999
Run-TryBot: Ian Lance Taylor <iant@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
This commit improves symbol table for XCOFF format.
It adds symbol alignment, TLS symbols and move the whole symbol table at
the end of the FILE. As relocations in the future external linking will
need symbols' index, we cannot write the symbol table when it's
generated.
Change-Id: I5dcae85b95e538b65f1a128faf56d4e2aa15baf1
Reviewed-on: https://go-review.googlesource.com/c/163998
Run-TryBot: Ian Lance Taylor <iant@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
This commit fixes a panic when -a was used with external linkmode.
Fixes#29807
Change-Id: I8cd42775f2953cec620cbc9ab345421c2694c9a2
Reviewed-on: https://go-review.googlesource.com/c/158998
Reviewed-by: Matthew Dempsky <mdempsky@google.com>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
This commit move data addresses to 0x200000000 for XCOFF executables.
.data and .bss must always be position-independent on AIX. This
modification allows to detect more easily if they aren't, as segfault
will be triggered.
Change-Id: Ied7a5b72b9f4ff9f870a1626cf07c48110635e62
Reviewed-on: https://go-review.googlesource.com/c/151040
Run-TryBot: Ian Lance Taylor <iant@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
This avoids a warning from old versions of the GNU linker or glibc.
No test because these old versions are not readily available.
I tested this by hand on CentOS 6.
Fixes#28722
Change-Id: I16640c9b83a79f759ec68fac64874803e74fbbfb
Reviewed-on: https://go-review.googlesource.com/c/153257
Run-TryBot: Ian Lance Taylor <iant@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
This commit changes the code generated for addressing symbols on AIX
operating system.
On AIX, every symbol accesses must be done via another symbol near the TOC,
named TOC anchor or TOC entry. This TOC anchor is a pointer to the symbol
address.
During Progedit function, when a symbol access is detected, its instructions
are modified to create a load on its TOC anchor and retrieve the symbol.
Change-Id: I00cf8f49c13004bc99fa8af13d549a709320f797
Reviewed-on: https://go-review.googlesource.com/c/151039
Run-TryBot: Ian Lance Taylor <iant@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
This commit fixes and improves the generation of dynamic symbols for
XCOFF files.
This mainly adds for every dynamic symbols a new symbol named
s.Extname().
Change-Id: I5b788f076d9a05e5d42f08eb1a74fd3e3efa9a86
Reviewed-on: https://go-review.googlesource.com/c/151038
Run-TryBot: Ian Lance Taylor <iant@golang.org>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
XCOFF files can't have multiples text or data sections. The name
of each type section must be .text, .data and .bss.
This commit also updates cmd/internal/objfile/xcoff.go to retrieve Go
sections using runtime symbols.
Change-Id: Ib6315f19dad2d154a4531fc6508e7cbd8bc94743
Reviewed-on: https://go-review.googlesource.com/c/151037
Run-TryBot: Ian Lance Taylor <iant@golang.org>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
On XCOFF, it is forbidden relocation of a DATA pointer to a text
section. It happens when a RODATA symbol needs a DATA symbol's address.
This commit moves every RODATA symbols with a R_ADDR on a data symbol to
.data sections to avoid these relocations.
Change-Id: I7f34d8e0ebdc8352a74e6b40e4c893d8d9419f4d
Reviewed-on: https://go-review.googlesource.com/c/146977
Reviewed-by: Ian Lance Taylor <iant@golang.org>
This commit adds AIX operating system to cmd/link package for ppc64
architecture.
Updates: #25893
Change-Id: I349e0a2658c31919b72117b62c4c9935c9af07c0
Reviewed-on: https://go-review.googlesource.com/c/138730
Run-TryBot: Ian Lance Taylor <iant@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
.rel symbol type is sym.SELFROSECT, and that makes .rel written
into .rdata section. But .rel stores code - jump table used for
external C functions. So we have to mark whole .rdata section
as executable (IMAGE_SCN_MEM_EXECUTE), because of .rel presence
in it.
Move .rel into .text section, and make .rdata section non executable.
I also had to move code that adjusted the size of .rel symbol
before calling textaddress, otherwise textaddress would not
calculate size of .text section correctly.
Fixes#25926
Change-Id: I4962f5de7b367410154c8709adfcd8472de9ac1a
Reviewed-on: https://go-review.googlesource.com/c/125455
Run-TryBot: Alex Brainman <alex.brainman@gmail.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
Splits part of dwarfgeneratedebugsyms into a new function,
dwarfGenerateDebugInfo which is called between deadcode elimination
and type name mangling.
This function takes care of collecting and processing the DIEs for
all functions and package-level variables and also generates DIEs
for all types used in the program.
Fixes#23733
Change-Id: I75ef0608fbed2dffc3be7a477f1b03e7e740ec61
Reviewed-on: https://go-review.googlesource.com/111237
Run-TryBot: Heschi Kreinick <heschi@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Heschi Kreinick <heschi@google.com>
The sym.Symbol 'Plt' and 'Got' field are used only with cgo and/or
external linking and are not needed for most symbols. Relocate them to
sym.AuxSymbol so as to shrink the main Symbol struct.
Updates #26186
Change-Id: I170d628a760be300a0c1f738f0998970e91ce3d6
Reviewed-on: https://go-review.googlesource.com/125478
Reviewed-by: Ian Lance Taylor <iant@golang.org>
This patch contains the remnants of CL (122482), which was intended to
reduce memory allocation in 'relocsym'. Another CL (113637) went in
first that included pretty much all of the code changes in 122482,
however there are some changes to comments that are worth preserving.
Change-Id: Iacdbd2bfe3b7ca2656596570f06ce9a646211913
Reviewed-on: https://go-review.googlesource.com/122482
Reviewed-by: Ian Lance Taylor <iant@golang.org>
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
Comment changes to fix typos, no code changes.
Change-Id: I6c915f183025587fc479d14f5d2c885767348b1b
Reviewed-on: https://go-review.googlesource.com/131615
Reviewed-by: Ian Lance Taylor <iant@golang.org>
Enable the Go linker to generate executables for windows/arm.
Generates PE relocation tables, which are used by Windows to
dynamically relocate the Go binary in memory. Windows on ARM
requires all modules to be relocatable, unlike x86/amd64 which are
permitted to have fixed base addresses.
Updates #26148
Change-Id: Ie63964ff52c2377e121b2885e9d05ec3ed8dc1cd
Reviewed-on: https://go-review.googlesource.com/125648
Run-TryBot: Ian Lance Taylor <iant@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
Copying sym.Reloc in loops hurts performance as
it has 48 byte size (on 64-bit platforms).
There are quite many symbols and each of them has more than 1
relocation (so, it's possible to have more than 1kk relocs).
The're also traversed more than once in some code paths.
By using pointers to them, copies are avoided.
For linking "hello world" example from net/http:
name old time/op new time/op delta
Linker-4 530ms ± 2% 521ms ± 3% -1.80% (p=0.000 n=17+20)
Change-Id: I6518aec69d6adcd137f84b5c089ceab4cb4ea2dd
Reviewed-on: https://go-review.googlesource.com/113636
Run-TryBot: Iskander Sharipov <iskander.sharipov@intel.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
Archreloc had this signature:
func(*Link, *sym.Reloc, *sym.Symbol, *int64) bool
The last *int64 argument is used as out parameter.
Passed valus could be allocated on stack, but escape analysis
fails here, leading to high number of unwanted allocs.
If instead 4th arg is passed by value, and modified values is returned,
no problems with allocations arise:
func(*Link, *sym.Reloc, *sym.Symbol, int64) (int64, bool)
There are 2 benefits:
1. code becomes more readable.
2. less allocations.
For linking "hello world" example from net/http:
name old time/op new time/op delta
Linker-4 530ms ± 2% 520ms ± 2% -1.83% (p=0.001 n=17+16)
It's top 1 in alloc_objects from memprofile:
flat flat% sum% cum cum%
229379 33.05% 33.05% 229379 33.05% cmd/link/internal/ld.relocsym
...
list relocsym:
229379 229379 (flat, cum) 33.05% of Total
229379 229379 183: var o int64
After the patch, ~230k of int64 allocs (~ 1.75mb) removed.
Passes toolshash-check (toolstash cmp).
Change-Id: I25504fe27967bcff70c4b7338790f3921d15473d
Reviewed-on: https://go-review.googlesource.com/113637
Run-TryBot: Iskander Sharipov <iskander.sharipov@intel.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
DWARF compression accounts for roughly 30% of the linker's time. This
CL switches from DefaultCompression to BestSpeed, which virtually
eliminates this time. This roughly halves the overhead of handling
DWARF in the linker:
name \ time/op nodwarf dwarf dwarf-speed
BuildCmdGoAll 10.0s ±11% 10.6s ± 5% 10.8s ± 5%
nodwarf +6.41% +8.03%
dwarf ~
LinkCmdGo 626ms ± 5% 1096ms ± 2% 860ms ± 2%
nodwarf +75.17% +37.36%
dwarf -21.59%
Previously, enabling DWARF had a 75% overhead in link time for cmd/go.
This change reduces this overhead to 37% (a 22% reduction).
The effect on binary size is minimal compared to DefaultCompression,
and still substantially better than no compression:
cmd/go bytes
nodwarf 10106953
dwarf 12159049 nodwarf+20%
dwarf-speed 12408905 nodwarf+23%
dwarf-nozlib 17766473 nodwarf+76%
Updates #26318.
Change-Id: I33bb7caa038a2753c29104501663daf4839e7054
Reviewed-on: https://go-review.googlesource.com/123356
Run-TryBot: Austin Clements <austin@google.com>
Reviewed-by: Heschi Kreinick <heschi@google.com>
Forked from CL 111895.
The trickiest part of this is that the binary layout code (blk,
elfshbits, and various other things) assumes a constant offset between
symbols' and sections' file locations and their virtual addresses.
Compression, of course, breaks this constant offset. But we need to
assign virtual addresses to everything before compression in order to
resolve relocations before compression. As a result, compression needs
to re-compute the "address" of the DWARF sections and symbols based on
their compressed size. Luckily, these are at the end of the file, so
this doesn't perturb any other sections or symbols. (And there is, of
course, a surprising amount of code that assumes the DWARF segment
comes last, so what's one more place?)
Relevant benchmarks:
name old time/op new time/op delta
StdCmd 10.3s ± 2% 10.8s ± 1% +5.43% (p=0.000 n=30+30)
name old text-bytes new text-bytes delta
HelloSize 746kB ± 0% 746kB ± 0% ~ (all equal)
CmdGoSize 8.41MB ± 0% 8.41MB ± 0% ~ (all equal)
[Geo mean] 2.50MB 2.50MB +0.00%
name old data-bytes new data-bytes delta
HelloSize 10.6kB ± 0% 10.6kB ± 0% ~ (all equal)
CmdGoSize 252kB ± 0% 252kB ± 0% ~ (all equal)
[Geo mean] 51.5kB 51.5kB +0.00%
name old bss-bytes new bss-bytes delta
HelloSize 125kB ± 0% 125kB ± 0% ~ (all equal)
CmdGoSize 145kB ± 0% 145kB ± 0% ~ (all equal)
[Geo mean] 135kB 135kB +0.00%
name old exe-bytes new exe-bytes delta
HelloSize 1.60MB ± 0% 1.05MB ± 0% -34.39% (p=0.000 n=30+30)
CmdGoSize 16.5MB ± 0% 11.3MB ± 0% -31.76% (p=0.000 n=30+30)
[Geo mean] 5.14MB 3.44MB -33.08%
Fixes#11799.
Updates #6853.
Change-Id: I64197afe4c01a237523a943088051ee056331c6f
Reviewed-on: https://go-review.googlesource.com/118276
Run-TryBot: Heschi Kreinick <heschi@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Austin Clements <austin@google.com>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
Currently these two forms of layout are done in a single pass. This
makes it difficult to compress DWARF sections because that must be
done after relocations are applied, which must happen after virtual
address layout, but we can't layout the file until we've compressed
the DWARF sections.
Fix this by separating the two layout steps. In the process, we can
also unify the copy-pasted code in Link.address to compute file
offsets, which currently has some unnecessary variation.
Unlike the current file offset computation, which depends on virtual
addresses, the new computation only uses file offsets and sizes. This
will let us compress the file representation of a segment and create
the file layout based on its on-disk size rather than its original
in-memory size.
Tested by comparing the test binary for the "strings" package on all
supported GOOS/GOARCH combinations. All binaries are identical
(except, of course, their build IDs).
This is a second attempt at CL 111682.
For #11799.
Fixes#25863.
Change-Id: If09f28771bb4d78dd392fd58b8d7c9d5f22b0b9e
Reviewed-on: https://go-review.googlesource.com/118716
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Heschi Kreinick <heschi@google.com>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
Currently these two forms of layout are done in a single pass. This
makes it difficult to compress DWARF sections because that must be
done after relocations are applied, which must happen after virtual
address layout, but we can't layout the file until we've compressed
the DWARF sections.
Fix this by separating the two layout steps. In the process, we can
also unify the copy-pasted code in Link.address to compute file
offsets. Currently, each instance of this is slightly different, but
there's no reason for it to be. For example, we don't perform
PEFILEALIGN alignment on Segrodata or Selreltodata even when HeadType
== Hwindows, but it turns out it doesn't matter whether you do or
don't because these segments simply don't exist on Windows. Hence, in
the unified code path, we do this alignment for all segments.
Likewise, there are two ways of computing Fileoff:
seg.Vaddr - prev.Vaddr + prev.Fileoff
and
prev.Fileoff + uint64(Rnd(int64(prev.Filelen), int64(*FlagRound)))
At the moment, these always have the same value, but the latter will
continue to work after we start compressing sections on disk.
Tested by comparing test binaries for all packages in std before and
after this change for GOOS={linux,windows,darwin,plan9}. All binaries
are identical.
For #11799.
Change-Id: If09f28771bb4d78dd392fd58b8d7c9d5f22b0b9f
Reviewed-on: https://go-review.googlesource.com/111682
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Heschi Kreinick <heschi@google.com>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
This is extension to https://golang.org/cl/113955 that handled
duplicated "unresolved relocation" errors.
For platforms with trampoline support, additional error is generated
per each undefined symbol. This breaks TestUndefinedRelocErrors test
on these platforms.
Proposed fix:
1. Changes error text to be identical to normal undefined reloc.
If relocation is undefined, jump to it will be unresolved
as well.
2. Introduces a map that can be used by all sites that
handle this kind of errors which makes it easier
to report such errors exactly once.
Errors on ppc64 before this change (note first 4 lines):
main.defined1: unresolved inter-package jump to main.undefined()
main.defined1: unresolved inter-package jump to main.undefined()
main.defined2: unresolved inter-package jump to main.undefined()
main.defined2: unresolved inter-package jump to main.undefined()
main.defined1: relocation target main.undefined not defined
main.defined2: relocation target main.undefined not defined
runtime.main_main·f: function main is undeclared in the main package
After this change:
main.defined1: relocation target main.undefined not defined
main.defined2: relocation target main.undefined not defined
runtime.main_main·f: function main is undeclared in the main package
Because of (1), errors output is the same on all platforms now.
Fixes#25753
Change-Id: Ic3084202a6fc5d4a6d2d0a93344f012b37fe58ed
Reviewed-on: https://go-review.googlesource.com/116676
Run-TryBot: Iskander Sharipov <iskander.sharipov@intel.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
For given program with 2 undefined relocations (main and undefined):
package main
func undefined()
func defined() int {
undefined()
undefined()
return 0
}
var x = defined()
"go tool link" produces these errors:
main.defined: relocation target main.undefined not defined
main.defined: relocation target main.undefined not defined
runtime.main_main·f: relocation target main.main not defined
main.defined: undefined: "main.undefined"
main.defined: undefined: "main.undefined"
runtime.main_main·f: undefined: "main.main"
After this CL is applied:
main.defined: relocation target main.undefined not defined
runtime.main_main·f: function main is undeclared in the main package
Fixes#10978
Improved error message for main proposed in #24809.
Change-Id: I4ba8547b1e143bbebeb4d6e29ea05d932124f037
Reviewed-on: https://go-review.googlesource.com/113955
Run-TryBot: Iskander Sharipov <iskander.sharipov@intel.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
Replace thread creation with calls to the pthread
library in libc.
Update #17490
Change-Id: I1e19965c45255deb849b059231252fc6a7861d6c
Reviewed-on: https://go-review.googlesource.com/108679
Run-TryBot: Keith Randall <khr@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
