cmd/compile: Devirtualize calls when concrete type behind interface is statically known

[navytux]

Please answer these questions before submitting your issue. Thanks!

What did you do?

Please consider the following program:

package xxx

type Iface interface {
        DoSomething()
}

type myStruct struct {
}

func (ms *myStruct) DoSomething() {
        println("mystruct·DoSomething")
}

func test() {
        var i Iface = &myStruct{}
        i.DoSomething()
}

(https://play.golang.org/p/1SSSkjvvcy)

Note in test(): type behind i is statically known to be *myStruct

What did you expect to see?

Code generated for test() directly calls myStruct.DoSomething

What did you see instead?

"".test t=1 size=80 args=0x0 locals=0x18
        0x0000 00000 (devirt.go:14)     TEXT    "".test(SB), $24-0
        0x0000 00000 (devirt.go:14)     MOVQ    (TLS), CX
        0x0009 00009 (devirt.go:14)     CMPQ    SP, 16(CX)
        0x000d 00013 (devirt.go:14)     JLS     73
        0x000f 00015 (devirt.go:14)     SUBQ    $24, SP
        0x0013 00019 (devirt.go:14)     MOVQ    BP, 16(SP)
        0x0018 00024 (devirt.go:14)     LEAQ    16(SP), BP
        0x001d 00029 (devirt.go:14)     FUNCDATA        $0, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
        0x001d 00029 (devirt.go:14)     FUNCDATA        $1, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
        0x001d 00029 (devirt.go:15)     LEAQ    type."".myStruct(SB), AX
        0x0024 00036 (devirt.go:15)     MOVQ    AX, (SP)
        0x0028 00040 (devirt.go:15)     PCDATA  $0, $0
        0x0028 00040 (devirt.go:15)     CALL    runtime.newobject(SB)
        0x002d 00045 (devirt.go:15)     MOVQ    8(SP), AX
        0x0032 00050 (devirt.go:16)     MOVQ    go.itab.*"".myStruct,"".Iface+32(SB), CX
        0x0039 00057 (devirt.go:16)     MOVQ    AX, (SP)
        0x003d 00061 (devirt.go:16)     PCDATA  $0, $0
        0x003d 00061 (devirt.go:16)     CALL    CX
        0x003f 00063 (devirt.go:17)     MOVQ    16(SP), BP
        0x0044 00068 (devirt.go:17)     ADDQ    $24, SP
        0x0048 00072 (devirt.go:17)     RET
        0x0049 00073 (devirt.go:17)     NOP
        0x0049 00073 (devirt.go:14)     PCDATA  $0, $-1
        0x0049 00073 (devirt.go:14)     CALL    runtime.morestack_noctxt(SB)
        0x004e 00078 (devirt.go:14)     JMP     0

Note the indirect call in offsets 00050 - 00061.

Does this issue reproduce with the latest release (go1.8)?

Yes

Context

This issue originally came in the context of using reflect.Typeof(v).Comparable() to detect whether a value can be used as key in a map:

kisielk/og-rek#30 (comment)

Current code for reflect.Typeof() is just a pointer cast to reflect/runtime.rtype + nil check and then the result is returned as reflect.Type interface:

https://github.com/golang/go/blob/f072283b/src/reflect/type.go#L1405
https://github.com/golang/go/blob/f072283b/src/reflect/type.go#L3025

So since result of reflect.Typeof(v) is statically known to be either:

• *reflect.rtype, or
• nil

code generated for reflect.Typeof(v).Comparable() should ideally be:

• cast v to rtype
• check for nil - if yes panic
• directly call rtype·Comparable()

but currently it does the call indirectly - similar to my above original example for myStruct:

...
        0x001d 00029 (x.go:6)   MOVQ    "".v+48(FP), AX
        0x0022 00034 (x.go:6)   MOVQ    AX, reflect.i·2+16(SP)
        0x0027 00039 (x.go:6)   MOVQ    "".v+56(FP), AX
        0x002c 00044 (x.go:6)   MOVQ    AX, reflect.i·2+24(SP)
        0x0031 00049 (x.go:6)   MOVQ    reflect.i·2+16(SP), AX
        0x0036 00054 (x.go:6)   TESTQ   AX, AX
        0x0039 00057 (x.go:6)   JEQ     95
        0x003b 00059 (x.go:6)   LEAQ    go.itab.*reflect.rtype,reflect.Type(SB), CX
        0x0042 00066 (x.go:6)   MOVQ    64(CX), CX
        0x0046 00070 (x.go:6)   MOVQ    AX, (SP)
        0x004a 00074 (x.go:6)   PCDATA  $0, $1
        0x004a 00074 (x.go:6)   CALL    CX
...

So implementing this kind of local devirtualization should imho help in many cases where reflect is used.

Possibly related issues:

#19165
#16869

System details

go version devel +f072283bce Thu Mar 2 06:08:42 2017 +0000 linux/amd64
GOARCH="amd64"
GOBIN=""
GOEXE=""
GOHOSTARCH="amd64"
GOHOSTOS="linux"
GOOS="linux"
GOPATH="/home/kirr/go"
GORACE=""
GOROOT="/home/kirr/src/tools/go/go"
GOTOOLDIR="/home/kirr/src/tools/go/go/pkg/tool/linux_amd64"
GCCGO="/usr/bin/gccgo"
CC="gcc"
GOGCCFLAGS="-fPIC -m64 -pthread -fmessage-length=0 -fdebug-prefix-map=/tmp/go-build669846645=/tmp/go-build -gno-record-gcc-switches"
CXX="g++"
CGO_ENABLED="1"
CGO_CFLAGS="-g -O2"
CGO_CPPFLAGS=""
CGO_CXXFLAGS="-g -O2"
CGO_FFLAGS="-g -O2"
CGO_LDFLAGS="-g -O2"
PKG_CONFIG="pkg-config"
GOROOT/bin/go version: go version devel +f072283bce Thu Mar 2 06:08:42 2017 +0000 linux/amd64
GOROOT/bin/go tool compile -V: compile version devel +f072283bce Thu Mar 2 06:08:42 2017 +0000 X:framepointer
uname -sr: Linux 4.9.0-1-amd64
Distributor ID:	Debian
Description:	Debian GNU/Linux 9.0 (stretch)
Release:	9.0
Codename:	stretch
/lib/x86_64-linux-gnu/libc.so.6: GNU C Library (Debian GLIBC 2.24-9) stable release version 2.24, by Roland McGrath et al.
gdb --version: GNU gdb (Debian 7.12-6) 7.12.0.20161007-git

Thanks beforehand,
Kirill

/cc @randall77, @dr2chase

[navytux]

I'd like to add: this can also help to reduce pressure on garbage collector.

Reasoning: currently when something is called via interface, both receiver and arguments are automatically escaped to heap, irregardless of what actually happens inside the call, because in general case when we do not know what code gets called, the worst needs to be assumed.

On the other hand when we could statically know the "underlying" type behind an interface, it is possible to use the information and ask escape analysis to do its work. As a consiquence in not so few cases things will stay on stack instead of being moved to heap.

Some examples:

  1 package xxx
  2 
  3 import "reflect"
  4 
  5 func comparable(v interface{}) bool {
  6         return reflect.TypeOf(v).Comparable()
  7 }       

devirt2.go:6: inlining call to reflect.TypeOf
devirt2.go:6: inlining call to reflect.toType
devirt2.go:6: reflect.Type(reflect.t·2) escapes to heap
devirt2.go:5: leaking param: v                                         <-- (at least this)
devirt2.go:6: comparable &reflect.i·2 does not escape

---

  1 package xxx
  2 
  3 import (
  4         "os"
  5         "io"
  6 )
  7 
  8 func readpipe() {
  9         r, w, err := os.Pipe()
 10         if err != nil {
 11                 panic(err)
 12         }
 13         _ = w
 14 
 15         rl := io.LimitedReader{R: r, N: 64}
 16         buf := [128]byte{}
 17 
 18         _, err = rl.Read(buf[:])
 19 }

devirt2.go:15: r escapes to heap        <--
devirt2.go:18: buf escapes to heap      <--
devirt2.go:16: moved to heap: buf       <--
devirt2.go:18: readpipe rl does not escape

---

etc.

This might be related to e.g. #18822 but not only on mutable/immutable but also on escapes/not escapes and similar.

So since garbage collection is known to have relatively high cost and statically tracking underlying interface types could reduce pressure to garbage collector, may I suggest to treat this issue as not low-priority?

Thanks again,
Kirill

[philhofer]

Part of the difficulty in covering many of the cases you've listed is that they would require the compiler to track values across call sites, and the compiler only performs intra-procedural optimizations today (perhaps with the exception of escape analysis). We'd have to make more aggressive inlining decisions to expose many of these opportunities. There's some work being done on the front, but it's a prerequisite to making this sort of optimization generally effective.

With that being said, in the very elementary case where you have

var r io.Reader
r = bytes.NewBuffer(nil)
r.Read(buf)

it should be possible to devirtualize the call to Read.

[navytux]

Yes, I understand. As you say we can start from something simple - your above example + cases where called something is already being inlined. E.g. reflect.TypeOf() is already being inlined so even this simple way would help imho.

[philhofer]

So, de-virtualizing the calls themselves is relatively straightforward: https://go-review.googlesource.com/c/37751/

We'll have to wait until inlining is performed post-SSA before this has an effect on some of your examples, but at least it gets the trivial ones.

[navytux]

Thanks a lot for starting this.

You mention there is some work being done also on inlining front. Any pointers? I'm not a compiler guy, but it is always interesting for me to look around and learn something.

Thanks beforehand,
Kirill

[gopherbot]

CL https://golang.org/cl/37751 mentions this issue.

[philhofer]

David's slides on inlining here: https://docs.google.com/presentation/d/1Wcblp3jpfeKwA0Y4FOmj63PW52M_qmNqlQkNaLj0P5o/edit#slide=id.p

[navytux]

Thanks for the link.

[gopherbot]

CL https://golang.org/cl/38139 mentions this issue.

[navytux]

@philhofer thanks for getting the first patch ready and merged. Looking forward for follow-up where situations with convT2I could be handled too. Hopefully reflect.Typeof() will be devertualized then.

Thanks beforehand,
Kirill

[randall77]

This is now all fixed at tip. We now devirtualize calls like this:

        var i Iface = &myStruct{}
        i.DoSomething()

and

	var r io.Reader
	r = bytes.NewBuffer(nil)
	r.Read(buf)

I'm sure there are more complicated cases we miss, but we get the examples in this issue.

[navytux]

@randall77, thanks for the update.