There is a small speedup:
(linking cmd/compile)
name old time/op new time/op delta
Deadcode 57.1ms ± 1% 53.5ms ± 1% -6.44% (p=0.008 n=5+5)
With this, we don't need a slice to read the relocations, reduce
some allocations.
name old alloc/op new alloc/op delta
Deadcode 4.16MB ± 0% 3.84MB ± 0% -7.85% (p=0.008 n=5+5)
Change-Id: Icd41c05682ba3f293a8cb9d2fe818e39d7276e5a
Reviewed-on: https://go-review.googlesource.com/c/go/+/222244
Reviewed-by: Than McIntosh <thanm@google.com>
Use a different mechanism to access relocations from the object
files, and use it in the stack bounds check pass. This shows some
speedup.
(linking cmd/compile)
Dostkcheck 76.9ms ± 1% 55.1ms ± 1% -28.36% (p=0.008 n=5+5)
Change-Id: I2ac42da515dccd64719fb557ffff6cdc69e4319b
Reviewed-on: https://go-review.googlesource.com/c/go/+/222240
Run-TryBot: Cherry Zhang <cherryyz@google.com>
Reviewed-by: Than McIntosh <thanm@google.com>
Move NoSplit flag from FuncInfo to symbol flag, so this can be
accessed easily without reading the FuncInfo.
The CFunc flag is never used. Remove.
Change-Id: I8bf4fcb2f209434bb90ccc4987a4c3f28f003323
Reviewed-on: https://go-review.googlesource.com/c/go/+/220058
Run-TryBot: Cherry Zhang <cherryyz@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Jeremy Faller <jeremy@golang.org>
Reviewed-by: Than McIntosh <thanm@google.com>
Most of the time of goobj2.Sym.Read is spent in reading the
symbol name. Sometimes the name is not needed, so don't read
it.
This brings the linking time pretty much back to the old linker.
On my Mac,
$ time go build cmd/compile # (new)
real 0m1.016s
user 0m1.164s
sys 0m0.297s
$ time go-tip build cmd/compile # (old)
real 0m0.980s
user 0m1.229s
sys 0m0.309s
Again, this is a bit ugly. Maybe we want to consider moving
symbol names out of Sym, to a separate section of the object
file?
Change-Id: Ia9085c15aa4d614b7eae126e71a79d6806c75055
Reviewed-on: https://go-review.googlesource.com/c/go/+/218480
Run-TryBot: Cherry Zhang <cherryyz@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Jeremy Faller <jeremy@golang.org>
Reviewed-by: Than McIntosh <thanm@google.com>
Reading unnecessary fields from object file take time, even from
memory mapping. In deadcode, we do this a alot, so only read
necessary ones. In particular, for relocations, for non-type
symbols, we only need their target symbols and reloc types to
build the dependency graph, so don't read other fields. This
should speed up the deadcode pass.
Also cache an object's number of defined package symbols, as it
is accessed a lot in resolve.
This is a bit ugly from an API point of view. It would be nice if
the compiler could figure out some fields are not used so don't
fill them.
Change-Id: I5c4e9526b8a3b0aead9fa71901a51fb214f013a7
Reviewed-on: https://go-review.googlesource.com/c/go/+/218479
Run-TryBot: Cherry Zhang <cherryyz@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Jeremy Faller <jeremy@golang.org>
Reviewed-by: Than McIntosh <thanm@google.com>
This change removes the NewReader function (no longer used by objdump)
and prunes away the now unused code paths from Reader.BytesAt and
Reader.StringAt, which helps with performance. At the moment the
reader operates by always ingesting the entire object file (either via
direct read or by mmap), meaning that there will always be a slice
available for us to index into.
Change-Id: I3af7396effe19e50ed594fe8d82fd2d15465687c
Reviewed-on: https://go-review.googlesource.com/c/go/+/201437
Run-TryBot: Than McIntosh <thanm@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Cherry Zhang <cherryyz@google.com>
Reviewed-by: Jeremy Faller <jeremy@golang.org>
Add a flag bit to mark symbols in the new object file as containing Go
type information. The use of a flag eliminates the need to do symbol
name matching as part of the new dead code elimination pass, which
should produce a minor speedup.
Change-Id: Iec8700e1139e2c4e310644c0766379865d2d6f82
Reviewed-on: https://go-review.googlesource.com/c/go/+/201399
Run-TryBot: Than McIntosh <thanm@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Jeremy Faller <jeremy@golang.org>
Reviewed-by: Cherry Zhang <cherryyz@google.com>
For the new object file format, don't tag individual symbols with a
"shared" flag, since that characteristic is better off as an attribute
of the containing object file as opposed to the individual symbol. Add
a new flags field in the object file header and put a bit in the flags
if the shared flags is in effect during compilation.
Change-Id: I2cf6d33bf7bf2fd8a7614ae0cd6ef03914777498
Reviewed-on: https://go-review.googlesource.com/c/go/+/201398
Run-TryBot: Than McIntosh <thanm@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Jeremy Faller <jeremy@golang.org>
Reviewed-by: Cherry Zhang <cherryyz@google.com>
Per Jeremy's comment in CL 199643. This makes the code read
better.
Change-Id: If270aecd712a27fb52e3faf5a4339200327d9ffe
Reviewed-on: https://go-review.googlesource.com/c/go/+/201023
Run-TryBot: Cherry Zhang <cherryyz@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Jeremy Faller <jeremy@golang.org>
Use the auxiliary symbol mechanism to connect the text symbol and
its associated DWARF symbols. This way, the linker can track the
DWARF symbols from the text symbol, without looking up special
names.
Currently, in the linker this is only used in the deadcode pass
to track which DWARF symbols are used and need to load. Later
passes still use name lookup for now.
Change-Id: I2fe49f3b1f0ecc1472ae8aa93907cff740022d8d
Reviewed-on: https://go-review.googlesource.com/c/go/+/199801
Run-TryBot: Cherry Zhang <cherryyz@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Than McIntosh <thanm@google.com>
In CL 196030 we decided to combine the imported package list
(Autolib) and referenced package list (PkgIdx, or Pkglist).
However, in some cases the Autolib list may contain file name,
instead of package path, e.g.
https://go.googlesource.com/go/+/refs/heads/dev.link/src/cmd/compile/internal/gc/main.go#1181
And the linker needs that to locate the file. This mostly happens
with direct invocation of the compiler and linker (i.e., not
through "go build").
Instead of letting the linker make guess of the file name based
on the package path, make Autolib a separate list.
Change-Id: If195a69462d04db515346ee67cdec925f5a69e2e
Reviewed-on: https://go-review.googlesource.com/c/go/+/200157
Run-TryBot: Cherry Zhang <cherryyz@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Jeremy Faller <jeremy@golang.org>
When reflect.Type.Method is called, all exported methods from a
reachable type need to be conservatively live. When such a
function is called, the compiler sets an attribute to the
function, and the linker needs to check that attribute. Implement
this in the index-based deadcode pass.
Unify symbol flags and FuncInfo flags to make things simpler. In
particular, the deadcode pass can check the reflectMethod
attribute without reading in and decoding FuncInfo.
Change-Id: Ibb21e172f2996e899c6efa5551a29d0eca62df67
Reviewed-on: https://go-review.googlesource.com/c/go/+/200099
Run-TryBot: Cherry Zhang <cherryyz@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Than McIntosh <thanm@google.com>
The dev.link branch was not sync'd with the new DWARF compilation
unit logic change on the master branch, and the new object file
format didn't support this.
This CL adds the new DWARF CU and file table support to the new
object file format. In the old object file, the DWARF file table
is a separate section. For now, we do the same with the new
object file, keeping it as a separate block.
While here, also refactor the loader code so it is easier for the
loader to carry per-object informations.
Change-Id: I4c317941fc0a5831acbc11ce8c2a8b7421471372
Reviewed-on: https://go-review.googlesource.com/c/go/+/198198
Reviewed-by: Austin Clements <austin@google.com>
With the old object file format, we use mmap (if supported) to
read object files and back symbol data with mapped read-only
memory.
Do the same with the new object file format. This also
significantly reduces number of syscalls made to read object
files.
Currently we still do mmap in object file level, not archive
level. This is probably ok, as there shouldn't be many archives
that contain more than one object. If this is a problem we can
change that later.
Change-Id: Icae3ef14d8ed6adbee1b5b48d420e2af22fd9604
Reviewed-on: https://go-review.googlesource.com/c/go/+/197797
Run-TryBot: Cherry Zhang <cherryyz@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Than McIntosh <thanm@google.com>
If -newobj is set, write object file in new format, which uses
indices for symbol references instead of symbol names. The file
format is described at the beginning of
cmd/internal/goobj2/objfile.go.
A new package, cmd/internal/goobj2, is introduced for reading and
writing new object files. (The package name is temporary.) It is
written in a way that trys to make the encoding as regular as
possible, and the reader and writer as symmetric as possible.
This is incomplete, and currently nothing will consume the new
object file.
Change-Id: Ifefedbf6456d760d15a9f40a28af6486c93100fe
Reviewed-on: https://go-review.googlesource.com/c/go/+/196030
Run-TryBot: Cherry Zhang <cherryyz@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Than McIntosh <thanm@google.com>
