cmd/link: dead code elimination for side-effect free functions #14840

mdempsky · 2016-03-16T20:36:38Z

Currently if we compile a Go program like:

package main
type dead int
func newDead() *dead { return new(dead) }
var x = newDead()
func main() {}

the linker can't dead code eliminate x or dead. This is because the "x = newDead()" initialization is compiled to an implicit init function, which causes the linker to pull in x and newDead.

(See https://golang.org/cl/20765 for a real world example.)

In general, just because x is otherwise unused, the linker can't get rid of the newDead() call because it might have side-effects. However, it should be possible for the compiler to help identify functions that are side-effect free, which could in turn let the linker be more aggressive about dead code elimination.

We would probably also need to tweak how cmd/compile generates package initializer functions for the linker to be able to eliminate individual initializers.

The text was updated successfully, but these errors were encountered:

bradfitz · 2016-03-16T20:48:54Z

This could help with shrinking init-time map literal construction in e.g. the unicode package.

If the unicode.init funcs generating the unicode's various maps were flagged as side-effect-free, then the whole init funcs can be deadcode eliminated if nobody referred to those maps.

/cc @crawshaw @josharian

randall77 · 2016-03-16T21:25:10Z

init methods are a problem because they may reference (& initialize)
otherwise dead globals. Anything we can do to initialize globals without
inits will help (see https://go-review.googlesource.com/c/17398/ for an
example). It doesn't look like this is an easy transformation here,
because newDead() could do anything.
If instead it was:

var x = new(dead)

maybe we could rewrite that in the compiler to

var _x dead
var x = &_x

the latter wouldn't need any init code, just relocations, so if x was
otherwise unused both x and _x would be stripped out at link time.

On Wed, Mar 16, 2016 at 1:49 PM, Brad Fitzpatrick notifications@github.com
wrote:

This could help with shrinking init-time map literal construction in e.g.
the unicode package.

If the unicode.init funcs generating the unicode's various maps were
flagged as side-effect-free, then the whole init funcs can be deadcode
eliminated if nobody referred to those maps.

/cc @crawshaw https://github.com/crawshaw @josharian
https://github.com/josharian

—
You are receiving this because you are subscribed to this thread.
Reply to this email directly or view it on GitHub
#14840 (comment)

mdempsky · 2016-03-16T21:34:12Z

@randall77 I think during escape analysis we could conservatively detect whether a function has side effects?

randall77 · 2016-03-16T21:53:22Z

I don't think the current escape analysis can help, as it runs after the init functions are built. You'd want to decide the side-effect-freeness of the RHS of global initializers before the init function is built.
Probably a very simple analysis would capture most of the cases we care about.

And then you'd have to have an init function per global, have it associated with the global, and only link it in if the global was otherwise reachable.

josharian · 2016-03-16T21:56:43Z

The compiler already has samesafeexpr. We could remove the "same" part and steal that.

I wonder whether have an init function per global (with all the associated overhead) would end up outweighing the other good done.

mdempsky · 2016-03-16T22:40:43Z

Overhead from too many init functions is a fair concern. I'd imagine a lot of the functions could be compiled as nosplit, which might reduce the overhead.

Also, traditionally ELF compilers generate .init sections with straight line instructions that are just pasted together, skipping function call overhead for simple functions. (The .init code still ends up forming a proper function because the linker will also include C runtime files like crti.o and crtn.o, which supply the appropriate function prologue/epilogue in their .init sections.)

crawshaw · 2016-03-17T13:47:34Z

Anyone writing code for this? If not, I'll spend a couple of days on it.

Assuming each global with side-effect free initialization is split off, I suspect the linker will need to merge the safe init functions together after reachability analysis. But that should be safe to do as long as we distinguish the safe generated init code from general init functions.

bradfitz · 2016-03-17T16:17:52Z

All yours.

mdempsky · 2016-03-17T18:04:27Z

@crawshaw Go for it.

crawshaw · 2016-03-21T13:49:51Z

I made a bit of progress on this, but after looking closer I don't know how it will help binary size.

The obvious target is the range tables in the unicode package. It would be nice to make it possible for the linker's deadcode pass to catch exported symbols like Nl, Nd, etc. But: all of these tables are included in the unicode.Categories map, which is used by the regexp package.

All the real world programs I looked at use package regexp. Even tiny ones like objdump. So while this will make the "hello world" program from #6853 look good, I don't see how it will help in practice.

If anyone can convince me otherwise I'll dust this off, but for now this looks like an academic exercise to me.

mdempsky · 2020-11-28T23:25:25Z

I'm going to look into extending escape analysis to keep track of side-effect free functions for Go 1.17.

ianlancetaylor · 2021-04-19T23:04:46Z

Not sure if this is happening, but moving milestone to backlog as it is a long-standing issue.

bradfitz · 2021-08-05T05:45:19Z

As a datapoint, I just tried changing the representation of unicode.RangeTable from a struct to a string containing the binary packed contents of the RangeTables.

Then instead of vars, I made them all consts. And instead of the five map[string]*RangeTable global variables, I made them funcs that returned a sync.Once-lazily-init map. (which the linker knows how to eliminate if unused)

I of course also updated all the funcs taking a *RangeTable pointer to instead just take a RangeTable string.

After minor corresponding updates to the regexp and encoding/xml packages (which both used *RangeTable), the binary size of a fmt-using Hello World program dropped 100 KB on Linux.

bradfitz@tsdev:~$ ls -l hello.before hello.after
-rwxr-xr-x 1 bradfitz bradfitz 1843533 Aug  4 22:20 hello.after
-rwxr-xr-x 1 bradfitz bradfitz 1942437 Aug  4 22:20 hello.before

bradfitz@tsdev:~$ size hello.before hello.after
   text    data     bss     dec     hex filename
1261840   88756  207456 1558052  17c624 hello.before
1234832   25752  207424 1468008  166668 hello.after

bradfitz@tsdev:~$ go tool nm hello.before | grep -F unicode. | wc -l
242
bradfitz@tsdev:~$ go tool nm hello.before | grep -F unicode. | head -20
  536f40 D unicode..inittask
  544318 D unicode.ASCII_Hex_Digit
  544320 D unicode.Adlam
  544328 D unicode.Ahom
  544330 D unicode.Anatolian_Hieroglyphs
  544338 D unicode.Arabic
  544340 D unicode.Armenian
  544348 D unicode.Avestan
  544350 D unicode.Balinese
  544358 D unicode.Bamum
  544360 D unicode.Bassa_Vah
  544368 D unicode.Batak
  544370 D unicode.Bengali
  544378 D unicode.Bhaiksuki
  544380 D unicode.Bidi_Control
  544388 D unicode.Bopomofo
  544390 D unicode.Brahmi
  544398 D unicode.Braille
  5443a0 D unicode.Buginese
  5443a8 D unicode.Buhid

bradfitz@tsdev:~$ go tool nm hello.after | grep -F unicode. | wc -l
1
bradfitz@tsdev:~$ go tool nm hello.after | grep -F unicode. 
  530280 D unicode..inittask

Which is all to say: there are wins to be had here if anybody is curious.

gopherbot · 2023-01-12T14:40:30Z

Change https://go.dev/cl/461315 mentions this issue: cmd/compile,cmd/link: enable deadcode of unreferenced large global maps

gopherbot · 2023-01-25T16:29:42Z

Change https://go.dev/cl/463395 mentions this issue: cmd/link: linker portion of dead map removal (patch 2 of 2)

gopherbot · 2023-01-27T13:18:14Z

Change https://go.dev/cl/463855 mentions this issue: cmd/internal/obj: flag init functions in object file

Introduce a flag in the object file indicating whether a given function corresponds to a compiler-generated (not user-written) init function, such as "os.init" or "syscall.init". Add code to the compiler to fill in the correct value for the flag, and add support to the loader package in the linker for testing the flag. The new loader API is currently unused, but will be needed in the next CL in this stack. Updates #2559. Updates #36021. Updates #14840. Change-Id: Iea7ad2adda487e4af7a44f062f9817977c53b394 Reviewed-on: https://go-review.googlesource.com/c/go/+/463855 Reviewed-by: Cherry Mui <cherryyz@google.com> Run-TryBot: Than McIntosh <thanm@google.com> TryBot-Result: Gopher Robot <gobot@golang.org>

This patch changes the compiler's pkg init machinery to pick out large initialization assignments to global maps (e.g. var mymap = map[string]int{"foo":1, "bar":2, ... } and extract the map init code into a separate outlined function, which is then called from the main init function with a weak relocation: var mymap map[string]int // KEEP reloc -> map.init.0 func init() { map.init.0() // weak relocation } func map.init.0() { mymap = map[string]int{"foo":1, "bar":2} } The map init outlining is done selectively (only in the case where the RHS code exceeds a size limit of 20 IR nodes). In order to ensure that a given map.init.NNN function is included when its corresponding map is live, we add dummy R_KEEP relocation from the map variable to the map init function. This first patch includes the main compiler compiler changes, and with the weak relocation addition disabled. Subsequent patch includes the requred linker changes along with switching to the call to the outlined routine to a weak relocation. See the later linker change for associated compile time performance numbers. Updates #2559. Updates #36021. Updates #14840. Change-Id: I1fd6fd6397772be1ebd3eb397caf68ae9a3147e9 Reviewed-on: https://go-review.googlesource.com/c/go/+/461315 Run-TryBot: Than McIntosh <thanm@google.com> Reviewed-by: Cherry Mui <cherryyz@google.com> TryBot-Result: Gopher Robot <gobot@golang.org>

This patch contains the linker changes needed to enable deadcoding of large unreferenced map variables, in combination with a previous compiler change. We add a new cleanup function that runs just after deadcode that looks for relocations in "init" funcs that are weak, of type R_CALL (and siblings), and are targeting an unreachable function. If we find such a relocation, after checking to make sure it targets a map.init.XXX helper, we redirect the relocation to a point to a no-op routine ("mapinitnoop") in the runtime. Compilebench results for this change: │ out.base.txt │ out.wrap.txt │ │ sec/op │ sec/op vs base │ Template 218.6m ± 2% 221.1m ± 1% ~ (p=0.129 n=39) Unicode 180.5m ± 1% 178.9m ± 1% -0.93% (p=0.006 n=39) GoTypes 1.162 ± 1% 1.156 ± 1% ~ (p=0.850 n=39) Compiler 143.6m ± 1% 142.6m ± 1% ~ (p=0.743 n=39) SSA 8.698 ± 1% 8.719 ± 1% ~ (p=0.145 n=39) Flate 142.6m ± 1% 143.9m ± 3% ~ (p=0.287 n=39) GoParser 247.7m ± 1% 248.8m ± 1% ~ (p=0.265 n=39) Reflect 588.0m ± 1% 590.4m ± 1% ~ (p=0.269 n=39) Tar 198.5m ± 1% 201.3m ± 1% +1.38% (p=0.005 n=39) XML 259.1m ± 1% 260.0m ± 1% ~ (p=0.376 n=39) LinkCompiler 746.8m ± 2% 747.8m ± 1% ~ (p=0.706 n=39) ExternalLinkCompiler 1.906 ± 0% 1.902 ± 1% ~ (p=0.207 n=40) LinkWithoutDebugCompiler 522.4m ± 21% 471.1m ± 1% -9.81% (p=0.000 n=40) StdCmd 21.32 ± 0% 21.39 ± 0% +0.32% (p=0.035 n=40) geomean 609.2m 606.0m -0.53% │ out.base.txt │ out.wrap.txt │ │ user-sec/op │ user-sec/op vs base │ Template 401.9m ± 3% 406.9m ± 2% ~ (p=0.291 n=39) Unicode 191.9m ± 6% 196.1m ± 3% ~ (p=0.052 n=39) GoTypes 3.967 ± 3% 4.056 ± 1% ~ (p=0.099 n=39) Compiler 171.1m ± 3% 173.4m ± 3% ~ (p=0.415 n=39) SSA 30.04 ± 4% 30.25 ± 4% ~ (p=0.106 n=39) Flate 246.3m ± 3% 248.9m ± 4% ~ (p=0.499 n=39) GoParser 518.7m ± 1% 520.6m ± 2% ~ (p=0.531 n=39) Reflect 1.670 ± 1% 1.656 ± 2% ~ (p=0.137 n=39) Tar 352.7m ± 2% 360.3m ± 2% ~ (p=0.117 n=39) XML 528.8m ± 2% 521.1m ± 2% ~ (p=0.296 n=39) LinkCompiler 1.128 ± 2% 1.140 ± 2% ~ (p=0.324 n=39) ExternalLinkCompiler 2.165 ± 2% 2.147 ± 2% ~ (p=0.537 n=40) LinkWithoutDebugCompiler 484.2m ± 4% 490.7m ± 3% ~ (p=0.897 n=40) geomean 818.5m 825.1m +0.80% │ out.base.txt │ out.wrap.txt │ │ text-bytes │ text-bytes vs base │ HelloSize 766.0Ki ± 0% 766.0Ki ± 0% ~ (p=1.000 n=40) ¹ CmdGoSize 10.02Mi ± 0% 10.02Mi ± 0% -0.03% (n=40) geomean 2.738Mi 2.738Mi -0.01% ¹ all samples are equal │ out.base.txt │ out.wrap.txt │ │ data-bytes │ data-bytes vs base │ HelloSize 14.17Ki ± 0% 14.17Ki ± 0% ~ (p=1.000 n=40) ¹ CmdGoSize 308.3Ki ± 0% 298.5Ki ± 0% -3.19% (n=40) geomean 66.10Ki 65.04Ki -1.61% ¹ all samples are equal │ out.base.txt │ out.wrap.txt │ │ bss-bytes │ bss-bytes vs base │ HelloSize 197.3Ki ± 0% 197.3Ki ± 0% ~ (p=1.000 n=40) ¹ CmdGoSize 228.2Ki ± 0% 228.1Ki ± 0% -0.01% (n=40) geomean 212.2Ki 212.1Ki -0.01% ¹ all samples are equal │ out.base.txt │ out.wrap.txt │ │ exe-bytes │ exe-bytes vs base │ HelloSize 1.192Mi ± 0% 1.192Mi ± 0% +0.00% (p=0.000 n=40) CmdGoSize 14.85Mi ± 0% 14.83Mi ± 0% -0.09% (n=40) geomean 4.207Mi 4.205Mi -0.05% Also tested for any linker changes by benchmarking relink of k8s "kubelet"; no changes to speak of there. Updates #2559. Updates #36021. Updates #14840. Change-Id: I4cc5370b3f20679a1065aaaf87bdf2881e257631 Reviewed-on: https://go-review.googlesource.com/c/go/+/463395 Run-TryBot: Than McIntosh <thanm@google.com> Reviewed-by: Cherry Mui <cherryyz@google.com> TryBot-Result: Gopher Robot <gobot@golang.org>

thanm · 2023-02-07T17:59:34Z

For posterity, linker dead code elimination of unused maps is now implemented in this stack.

Getting back to the original example:

package main
type dead int
func newDead() *dead { return new(dead) }
var x = newDead()
func main() {}

In principle the same strategy used for maps could be used here for variables like "x", e.g. outline the code that performs the init (e.g. call to newDead), then have a weak relocation from main.init to the outlined init func, and a strong relocation from "x" to the outlined init func.

The existing code that checks for side effects is here; in theory if were were to extend this code to work interprocedurally we might be able to catch this case.

Introduce a flag in the object file indicating whether a given function corresponds to a compiler-generated (not user-written) init function, such as "os.init" or "syscall.init". Add code to the compiler to fill in the correct value for the flag, and add support to the loader package in the linker for testing the flag. The new loader API is currently unused, but will be needed in the next CL in this stack. Updates golang#2559. Updates golang#36021. Updates golang#14840. Change-Id: Iea7ad2adda487e4af7a44f062f9817977c53b394 Reviewed-on: https://go-review.googlesource.com/c/go/+/463855 Reviewed-by: Cherry Mui <cherryyz@google.com> Run-TryBot: Than McIntosh <thanm@google.com> TryBot-Result: Gopher Robot <gobot@golang.org>

This patch changes the compiler's pkg init machinery to pick out large initialization assignments to global maps (e.g. var mymap = map[string]int{"foo":1, "bar":2, ... } and extract the map init code into a separate outlined function, which is then called from the main init function with a weak relocation: var mymap map[string]int // KEEP reloc -> map.init.0 func init() { map.init.0() // weak relocation } func map.init.0() { mymap = map[string]int{"foo":1, "bar":2} } The map init outlining is done selectively (only in the case where the RHS code exceeds a size limit of 20 IR nodes). In order to ensure that a given map.init.NNN function is included when its corresponding map is live, we add dummy R_KEEP relocation from the map variable to the map init function. This first patch includes the main compiler compiler changes, and with the weak relocation addition disabled. Subsequent patch includes the requred linker changes along with switching to the call to the outlined routine to a weak relocation. See the later linker change for associated compile time performance numbers. Updates golang#2559. Updates golang#36021. Updates golang#14840. Change-Id: I1fd6fd6397772be1ebd3eb397caf68ae9a3147e9 Reviewed-on: https://go-review.googlesource.com/c/go/+/461315 Run-TryBot: Than McIntosh <thanm@google.com> Reviewed-by: Cherry Mui <cherryyz@google.com> TryBot-Result: Gopher Robot <gobot@golang.org>

This patch contains the linker changes needed to enable deadcoding of large unreferenced map variables, in combination with a previous compiler change. We add a new cleanup function that runs just after deadcode that looks for relocations in "init" funcs that are weak, of type R_CALL (and siblings), and are targeting an unreachable function. If we find such a relocation, after checking to make sure it targets a map.init.XXX helper, we redirect the relocation to a point to a no-op routine ("mapinitnoop") in the runtime. Compilebench results for this change: │ out.base.txt │ out.wrap.txt │ │ sec/op │ sec/op vs base │ Template 218.6m ± 2% 221.1m ± 1% ~ (p=0.129 n=39) Unicode 180.5m ± 1% 178.9m ± 1% -0.93% (p=0.006 n=39) GoTypes 1.162 ± 1% 1.156 ± 1% ~ (p=0.850 n=39) Compiler 143.6m ± 1% 142.6m ± 1% ~ (p=0.743 n=39) SSA 8.698 ± 1% 8.719 ± 1% ~ (p=0.145 n=39) Flate 142.6m ± 1% 143.9m ± 3% ~ (p=0.287 n=39) GoParser 247.7m ± 1% 248.8m ± 1% ~ (p=0.265 n=39) Reflect 588.0m ± 1% 590.4m ± 1% ~ (p=0.269 n=39) Tar 198.5m ± 1% 201.3m ± 1% +1.38% (p=0.005 n=39) XML 259.1m ± 1% 260.0m ± 1% ~ (p=0.376 n=39) LinkCompiler 746.8m ± 2% 747.8m ± 1% ~ (p=0.706 n=39) ExternalLinkCompiler 1.906 ± 0% 1.902 ± 1% ~ (p=0.207 n=40) LinkWithoutDebugCompiler 522.4m ± 21% 471.1m ± 1% -9.81% (p=0.000 n=40) StdCmd 21.32 ± 0% 21.39 ± 0% +0.32% (p=0.035 n=40) geomean 609.2m 606.0m -0.53% │ out.base.txt │ out.wrap.txt │ │ user-sec/op │ user-sec/op vs base │ Template 401.9m ± 3% 406.9m ± 2% ~ (p=0.291 n=39) Unicode 191.9m ± 6% 196.1m ± 3% ~ (p=0.052 n=39) GoTypes 3.967 ± 3% 4.056 ± 1% ~ (p=0.099 n=39) Compiler 171.1m ± 3% 173.4m ± 3% ~ (p=0.415 n=39) SSA 30.04 ± 4% 30.25 ± 4% ~ (p=0.106 n=39) Flate 246.3m ± 3% 248.9m ± 4% ~ (p=0.499 n=39) GoParser 518.7m ± 1% 520.6m ± 2% ~ (p=0.531 n=39) Reflect 1.670 ± 1% 1.656 ± 2% ~ (p=0.137 n=39) Tar 352.7m ± 2% 360.3m ± 2% ~ (p=0.117 n=39) XML 528.8m ± 2% 521.1m ± 2% ~ (p=0.296 n=39) LinkCompiler 1.128 ± 2% 1.140 ± 2% ~ (p=0.324 n=39) ExternalLinkCompiler 2.165 ± 2% 2.147 ± 2% ~ (p=0.537 n=40) LinkWithoutDebugCompiler 484.2m ± 4% 490.7m ± 3% ~ (p=0.897 n=40) geomean 818.5m 825.1m +0.80% │ out.base.txt │ out.wrap.txt │ │ text-bytes │ text-bytes vs base │ HelloSize 766.0Ki ± 0% 766.0Ki ± 0% ~ (p=1.000 n=40) ¹ CmdGoSize 10.02Mi ± 0% 10.02Mi ± 0% -0.03% (n=40) geomean 2.738Mi 2.738Mi -0.01% ¹ all samples are equal │ out.base.txt │ out.wrap.txt │ │ data-bytes │ data-bytes vs base │ HelloSize 14.17Ki ± 0% 14.17Ki ± 0% ~ (p=1.000 n=40) ¹ CmdGoSize 308.3Ki ± 0% 298.5Ki ± 0% -3.19% (n=40) geomean 66.10Ki 65.04Ki -1.61% ¹ all samples are equal │ out.base.txt │ out.wrap.txt │ │ bss-bytes │ bss-bytes vs base │ HelloSize 197.3Ki ± 0% 197.3Ki ± 0% ~ (p=1.000 n=40) ¹ CmdGoSize 228.2Ki ± 0% 228.1Ki ± 0% -0.01% (n=40) geomean 212.2Ki 212.1Ki -0.01% ¹ all samples are equal │ out.base.txt │ out.wrap.txt │ │ exe-bytes │ exe-bytes vs base │ HelloSize 1.192Mi ± 0% 1.192Mi ± 0% +0.00% (p=0.000 n=40) CmdGoSize 14.85Mi ± 0% 14.83Mi ± 0% -0.09% (n=40) geomean 4.207Mi 4.205Mi -0.05% Also tested for any linker changes by benchmarking relink of k8s "kubelet"; no changes to speak of there. Updates golang#2559. Updates golang#36021. Updates golang#14840. Change-Id: I4cc5370b3f20679a1065aaaf87bdf2881e257631 Reviewed-on: https://go-review.googlesource.com/c/go/+/463395 Run-TryBot: Than McIntosh <thanm@google.com> Reviewed-by: Cherry Mui <cherryyz@google.com> TryBot-Result: Gopher Robot <gobot@golang.org>

This patch changes the compiler's pkg init machinery to pick out large initialization assignments to global maps (e.g. var mymap = map[string]int{"foo":1, "bar":2, ... } and extract the map init code into a separate outlined function, which is then called from the main init function with a weak relocation: var mymap map[string]int // KEEP reloc -> map.init.0 func init() { map.init.0() // weak relocation } func map.init.0() { mymap = map[string]int{"foo":1, "bar":2} } The map init outlining is done selectively (only in the case where the RHS code exceeds a size limit of 20 IR nodes). In order to ensure that a given map.init.NNN function is included when its corresponding map is live, we add dummy R_KEEP relocation from the map variable to the map init function. This first patch includes the main compiler compiler changes, and with the weak relocation addition disabled. Subsequent patch includes the requred linker changes along with switching to the call to the outlined routine to a weak relocation. See the later linker change for associated compile time performance numbers. Updates golang#2559. Updates golang#36021. Updates golang#14840. Change-Id: I1fd6fd6397772be1ebd3eb397caf68ae9a3147e9 Reviewed-on: https://go-review.googlesource.com/c/go/+/461315 Run-TryBot: Than McIntosh <thanm@google.com> Reviewed-by: Cherry Mui <cherryyz@google.com> TryBot-Result: Gopher Robot <gobot@golang.org>

This patch contains the linker changes needed to enable deadcoding of large unreferenced map variables, in combination with a previous compiler change. We add a new cleanup function that runs just after deadcode that looks for relocations in "init" funcs that are weak, of type R_CALL (and siblings), and are targeting an unreachable function. If we find such a relocation, after checking to make sure it targets a map.init.XXX helper, we redirect the relocation to a point to a no-op routine ("mapinitnoop") in the runtime. Compilebench results for this change: │ out.base.txt │ out.wrap.txt │ │ sec/op │ sec/op vs base │ Template 218.6m ± 2% 221.1m ± 1% ~ (p=0.129 n=39) Unicode 180.5m ± 1% 178.9m ± 1% -0.93% (p=0.006 n=39) GoTypes 1.162 ± 1% 1.156 ± 1% ~ (p=0.850 n=39) Compiler 143.6m ± 1% 142.6m ± 1% ~ (p=0.743 n=39) SSA 8.698 ± 1% 8.719 ± 1% ~ (p=0.145 n=39) Flate 142.6m ± 1% 143.9m ± 3% ~ (p=0.287 n=39) GoParser 247.7m ± 1% 248.8m ± 1% ~ (p=0.265 n=39) Reflect 588.0m ± 1% 590.4m ± 1% ~ (p=0.269 n=39) Tar 198.5m ± 1% 201.3m ± 1% +1.38% (p=0.005 n=39) XML 259.1m ± 1% 260.0m ± 1% ~ (p=0.376 n=39) LinkCompiler 746.8m ± 2% 747.8m ± 1% ~ (p=0.706 n=39) ExternalLinkCompiler 1.906 ± 0% 1.902 ± 1% ~ (p=0.207 n=40) LinkWithoutDebugCompiler 522.4m ± 21% 471.1m ± 1% -9.81% (p=0.000 n=40) StdCmd 21.32 ± 0% 21.39 ± 0% +0.32% (p=0.035 n=40) geomean 609.2m 606.0m -0.53% │ out.base.txt │ out.wrap.txt │ │ user-sec/op │ user-sec/op vs base │ Template 401.9m ± 3% 406.9m ± 2% ~ (p=0.291 n=39) Unicode 191.9m ± 6% 196.1m ± 3% ~ (p=0.052 n=39) GoTypes 3.967 ± 3% 4.056 ± 1% ~ (p=0.099 n=39) Compiler 171.1m ± 3% 173.4m ± 3% ~ (p=0.415 n=39) SSA 30.04 ± 4% 30.25 ± 4% ~ (p=0.106 n=39) Flate 246.3m ± 3% 248.9m ± 4% ~ (p=0.499 n=39) GoParser 518.7m ± 1% 520.6m ± 2% ~ (p=0.531 n=39) Reflect 1.670 ± 1% 1.656 ± 2% ~ (p=0.137 n=39) Tar 352.7m ± 2% 360.3m ± 2% ~ (p=0.117 n=39) XML 528.8m ± 2% 521.1m ± 2% ~ (p=0.296 n=39) LinkCompiler 1.128 ± 2% 1.140 ± 2% ~ (p=0.324 n=39) ExternalLinkCompiler 2.165 ± 2% 2.147 ± 2% ~ (p=0.537 n=40) LinkWithoutDebugCompiler 484.2m ± 4% 490.7m ± 3% ~ (p=0.897 n=40) geomean 818.5m 825.1m +0.80% │ out.base.txt │ out.wrap.txt │ │ text-bytes │ text-bytes vs base │ HelloSize 766.0Ki ± 0% 766.0Ki ± 0% ~ (p=1.000 n=40) ¹ CmdGoSize 10.02Mi ± 0% 10.02Mi ± 0% -0.03% (n=40) geomean 2.738Mi 2.738Mi -0.01% ¹ all samples are equal │ out.base.txt │ out.wrap.txt │ │ data-bytes │ data-bytes vs base │ HelloSize 14.17Ki ± 0% 14.17Ki ± 0% ~ (p=1.000 n=40) ¹ CmdGoSize 308.3Ki ± 0% 298.5Ki ± 0% -3.19% (n=40) geomean 66.10Ki 65.04Ki -1.61% ¹ all samples are equal │ out.base.txt │ out.wrap.txt │ │ bss-bytes │ bss-bytes vs base │ HelloSize 197.3Ki ± 0% 197.3Ki ± 0% ~ (p=1.000 n=40) ¹ CmdGoSize 228.2Ki ± 0% 228.1Ki ± 0% -0.01% (n=40) geomean 212.2Ki 212.1Ki -0.01% ¹ all samples are equal │ out.base.txt │ out.wrap.txt │ │ exe-bytes │ exe-bytes vs base │ HelloSize 1.192Mi ± 0% 1.192Mi ± 0% +0.00% (p=0.000 n=40) CmdGoSize 14.85Mi ± 0% 14.83Mi ± 0% -0.09% (n=40) geomean 4.207Mi 4.205Mi -0.05% Also tested for any linker changes by benchmarking relink of k8s "kubelet"; no changes to speak of there. Updates golang#2559. Updates golang#36021. Updates golang#14840. Change-Id: I4cc5370b3f20679a1065aaaf87bdf2881e257631 Reviewed-on: https://go-review.googlesource.com/c/go/+/463395 Run-TryBot: Than McIntosh <thanm@google.com> Reviewed-by: Cherry Mui <cherryyz@google.com> TryBot-Result: Gopher Robot <gobot@golang.org>

Introduce a flag in the object file indicating whether a given function corresponds to a compiler-generated (not user-written) init function, such as "os.init" or "syscall.init". Add code to the compiler to fill in the correct value for the flag, and add support to the loader package in the linker for testing the flag. The new loader API is currently unused, but will be needed in the next CL in this stack. Updates golang#2559. Updates golang#36021. Updates golang#14840. Change-Id: Iea7ad2adda487e4af7a44f062f9817977c53b394 Reviewed-on: https://go-review.googlesource.com/c/go/+/463855 Reviewed-by: Cherry Mui <cherryyz@google.com> Run-TryBot: Than McIntosh <thanm@google.com> TryBot-Result: Gopher Robot <gobot@golang.org>

This patch changes the compiler's pkg init machinery to pick out large initialization assignments to global maps (e.g. var mymap = map[string]int{"foo":1, "bar":2, ... } and extract the map init code into a separate outlined function, which is then called from the main init function with a weak relocation: var mymap map[string]int // KEEP reloc -> map.init.0 func init() { map.init.0() // weak relocation } func map.init.0() { mymap = map[string]int{"foo":1, "bar":2} } The map init outlining is done selectively (only in the case where the RHS code exceeds a size limit of 20 IR nodes). In order to ensure that a given map.init.NNN function is included when its corresponding map is live, we add dummy R_KEEP relocation from the map variable to the map init function. This first patch includes the main compiler compiler changes, and with the weak relocation addition disabled. Subsequent patch includes the requred linker changes along with switching to the call to the outlined routine to a weak relocation. See the later linker change for associated compile time performance numbers. Updates golang#2559. Updates golang#36021. Updates golang#14840. Change-Id: I1fd6fd6397772be1ebd3eb397caf68ae9a3147e9 Reviewed-on: https://go-review.googlesource.com/c/go/+/461315 Run-TryBot: Than McIntosh <thanm@google.com> Reviewed-by: Cherry Mui <cherryyz@google.com> TryBot-Result: Gopher Robot <gobot@golang.org>

This patch contains the linker changes needed to enable deadcoding of large unreferenced map variables, in combination with a previous compiler change. We add a new cleanup function that runs just after deadcode that looks for relocations in "init" funcs that are weak, of type R_CALL (and siblings), and are targeting an unreachable function. If we find such a relocation, after checking to make sure it targets a map.init.XXX helper, we redirect the relocation to a point to a no-op routine ("mapinitnoop") in the runtime. Compilebench results for this change: │ out.base.txt │ out.wrap.txt │ │ sec/op │ sec/op vs base │ Template 218.6m ± 2% 221.1m ± 1% ~ (p=0.129 n=39) Unicode 180.5m ± 1% 178.9m ± 1% -0.93% (p=0.006 n=39) GoTypes 1.162 ± 1% 1.156 ± 1% ~ (p=0.850 n=39) Compiler 143.6m ± 1% 142.6m ± 1% ~ (p=0.743 n=39) SSA 8.698 ± 1% 8.719 ± 1% ~ (p=0.145 n=39) Flate 142.6m ± 1% 143.9m ± 3% ~ (p=0.287 n=39) GoParser 247.7m ± 1% 248.8m ± 1% ~ (p=0.265 n=39) Reflect 588.0m ± 1% 590.4m ± 1% ~ (p=0.269 n=39) Tar 198.5m ± 1% 201.3m ± 1% +1.38% (p=0.005 n=39) XML 259.1m ± 1% 260.0m ± 1% ~ (p=0.376 n=39) LinkCompiler 746.8m ± 2% 747.8m ± 1% ~ (p=0.706 n=39) ExternalLinkCompiler 1.906 ± 0% 1.902 ± 1% ~ (p=0.207 n=40) LinkWithoutDebugCompiler 522.4m ± 21% 471.1m ± 1% -9.81% (p=0.000 n=40) StdCmd 21.32 ± 0% 21.39 ± 0% +0.32% (p=0.035 n=40) geomean 609.2m 606.0m -0.53% │ out.base.txt │ out.wrap.txt │ │ user-sec/op │ user-sec/op vs base │ Template 401.9m ± 3% 406.9m ± 2% ~ (p=0.291 n=39) Unicode 191.9m ± 6% 196.1m ± 3% ~ (p=0.052 n=39) GoTypes 3.967 ± 3% 4.056 ± 1% ~ (p=0.099 n=39) Compiler 171.1m ± 3% 173.4m ± 3% ~ (p=0.415 n=39) SSA 30.04 ± 4% 30.25 ± 4% ~ (p=0.106 n=39) Flate 246.3m ± 3% 248.9m ± 4% ~ (p=0.499 n=39) GoParser 518.7m ± 1% 520.6m ± 2% ~ (p=0.531 n=39) Reflect 1.670 ± 1% 1.656 ± 2% ~ (p=0.137 n=39) Tar 352.7m ± 2% 360.3m ± 2% ~ (p=0.117 n=39) XML 528.8m ± 2% 521.1m ± 2% ~ (p=0.296 n=39) LinkCompiler 1.128 ± 2% 1.140 ± 2% ~ (p=0.324 n=39) ExternalLinkCompiler 2.165 ± 2% 2.147 ± 2% ~ (p=0.537 n=40) LinkWithoutDebugCompiler 484.2m ± 4% 490.7m ± 3% ~ (p=0.897 n=40) geomean 818.5m 825.1m +0.80% │ out.base.txt │ out.wrap.txt │ │ text-bytes │ text-bytes vs base │ HelloSize 766.0Ki ± 0% 766.0Ki ± 0% ~ (p=1.000 n=40) ¹ CmdGoSize 10.02Mi ± 0% 10.02Mi ± 0% -0.03% (n=40) geomean 2.738Mi 2.738Mi -0.01% ¹ all samples are equal │ out.base.txt │ out.wrap.txt │ │ data-bytes │ data-bytes vs base │ HelloSize 14.17Ki ± 0% 14.17Ki ± 0% ~ (p=1.000 n=40) ¹ CmdGoSize 308.3Ki ± 0% 298.5Ki ± 0% -3.19% (n=40) geomean 66.10Ki 65.04Ki -1.61% ¹ all samples are equal │ out.base.txt │ out.wrap.txt │ │ bss-bytes │ bss-bytes vs base │ HelloSize 197.3Ki ± 0% 197.3Ki ± 0% ~ (p=1.000 n=40) ¹ CmdGoSize 228.2Ki ± 0% 228.1Ki ± 0% -0.01% (n=40) geomean 212.2Ki 212.1Ki -0.01% ¹ all samples are equal │ out.base.txt │ out.wrap.txt │ │ exe-bytes │ exe-bytes vs base │ HelloSize 1.192Mi ± 0% 1.192Mi ± 0% +0.00% (p=0.000 n=40) CmdGoSize 14.85Mi ± 0% 14.83Mi ± 0% -0.09% (n=40) geomean 4.207Mi 4.205Mi -0.05% Also tested for any linker changes by benchmarking relink of k8s "kubelet"; no changes to speak of there. Updates golang#2559. Updates golang#36021. Updates golang#14840. Change-Id: I4cc5370b3f20679a1065aaaf87bdf2881e257631 Reviewed-on: https://go-review.googlesource.com/c/go/+/463395 Run-TryBot: Than McIntosh <thanm@google.com> Reviewed-by: Cherry Mui <cherryyz@google.com> TryBot-Result: Gopher Robot <gobot@golang.org>

mdempsky added this to the Unplanned milestone Mar 16, 2016

chewxy mentioned this issue Feb 26, 2018

cmd/compile: enable mid-stack inlining #19348

Closed

This was referenced Jul 7, 2018

cmd/link: linker should be able to remove init functions with no side-effects #19533

Open

encoding/base64: make encoding/base64.init() smaller and faster #22450

Open

aclements mentioned this issue Nov 25, 2018

cmd/compile: static init maps should never generate code #2559

Open

rasky mentioned this issue Apr 11, 2020

x/net/idna: adds 680KiB to binary when imported but unreferenced #38311

Open

bradfitz mentioned this issue Apr 15, 2020

proposal: spec: lazy init imports to possibly import without side effects #38450

Closed

bradfitz added the binary-size label May 1, 2020

mdempsky modified the milestones: Unplanned, Go1.17 Nov 28, 2020

mdempsky self-assigned this Nov 28, 2020

ianlancetaylor added the NeedsInvestigation Someone must examine and confirm this is a valid issue and not a duplicate of an existing one. label Apr 19, 2021

ianlancetaylor modified the milestones: Go1.17, Backlog Apr 19, 2021

gopherbot added the compiler/runtime Issues related to the Go compiler and/or runtime. label Jul 13, 2022

dsnet mentioned this issue Aug 22, 2023

cmd/compile: dead code-elimination should work for sync.OnceValue #62202

Open

steiza mentioned this issue Nov 3, 2023

Reduce dependencies sigstore/sigstore-go#23

Open

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

cmd/link: dead code elimination for side-effect free functions #14840

cmd/link: dead code elimination for side-effect free functions #14840

mdempsky commented Mar 16, 2016

bradfitz commented Mar 16, 2016

randall77 commented Mar 16, 2016

mdempsky commented Mar 16, 2016

randall77 commented Mar 16, 2016

josharian commented Mar 16, 2016

mdempsky commented Mar 16, 2016

crawshaw commented Mar 17, 2016

bradfitz commented Mar 17, 2016

mdempsky commented Mar 17, 2016

crawshaw commented Mar 21, 2016

mdempsky commented Nov 28, 2020

ianlancetaylor commented Apr 19, 2021

bradfitz commented Aug 5, 2021

gopherbot commented Jan 12, 2023

gopherbot commented Jan 25, 2023

gopherbot commented Jan 27, 2023

thanm commented Feb 7, 2023

cmd/link: dead code elimination for side-effect free functions #14840

cmd/link: dead code elimination for side-effect free functions #14840

Comments

mdempsky commented Mar 16, 2016

bradfitz commented Mar 16, 2016

randall77 commented Mar 16, 2016

mdempsky commented Mar 16, 2016

randall77 commented Mar 16, 2016

josharian commented Mar 16, 2016

mdempsky commented Mar 16, 2016

crawshaw commented Mar 17, 2016

bradfitz commented Mar 17, 2016

mdempsky commented Mar 17, 2016

crawshaw commented Mar 21, 2016

mdempsky commented Nov 28, 2020

ianlancetaylor commented Apr 19, 2021

bradfitz commented Aug 5, 2021

gopherbot commented Jan 12, 2023

gopherbot commented Jan 25, 2023

gopherbot commented Jan 27, 2023

thanm commented Feb 7, 2023