cmd/compile: DUFFZERO in autogenerated methods corrupts LR

[davecheney]

The following program faults on ppc64le

http://play.golang.org/p/hRlQs7hss2

package main

type APIEndpoint struct {
        Addresses []string
        Hostnames []string
        CACert string
        EnvironUUID string
        ServerUUID string
}

type Base struct {
        name string
}

func (b *Base) ConnectionEndpoint() (APIEndpoint, error) {
        return APIEndpoint{}, nil
}

type Super struct {
        Base
}

type T interface {
    ConnectionEndpoint() (APIEndpoint, error)
}

func main() {
    var q T = new(Super)
    q.ConnectionEndpoint()
}

% go run x.go 
panic: runtime error: invalid memory address or nil pointer dereference
[signal 0xb code=0x1 addr=0x0 pc=0x11208]

goroutine 1 [running]:
main.(*Super).ConnectionEndpoint(0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, ...)
        <autogenerated>:1 +0x48
main.main()
        /home/ubuntu/src/x/x.go:29 +0x9c
exit status 2

The program does not fault on any other architecture.

/cc @rsc @aclements @minux

[davecheney]

Reducing the size of APIEndpoint by arbitrarily commenting out fields makes the test pass once the struct is 8 words of less.

[davecheney]

Here is a shorter example

package main

const N = 9 // values > 8 cause ppc64 to fault

type Base struct {
    name string
}

func (b *Base) ConnectionEndpoint() (x [N]uintptr, e error) {
    return
}

type Super struct {
    Base
}

type T interface {
    ConnectionEndpoint() ([N]uintptr, error)
}

func f(q T) {
    q.ConnectionEndpoint()
}

func main() {
    var s Super
    f(&s)
}

http://play.golang.org/p/PWbXrURQOD

[davecheney]

It's duffzero

Program received signal SIGSEGV, Segmentation fault.
0x000000000001121c in main.(*Super).ConnectionEndpoint (this=0x0, x=...) at <autogenerated>:2
2       <autogenerated>: No such file or directory.
(gdb) disassemble 
Dump of assembler code for function main.(*Super).ConnectionEndpoint:
   0x00000000000111d0 <+0>:     ld      r3,16(r30)
   0x00000000000111d4 <+4>:     cmpld   r3,r1
   0x00000000000111d8 <+8>:     blt     0x111e8 <main.(*Super).ConnectionEndpoint+24>
   0x00000000000111dc <+12>:    mflr    r5
   0x00000000000111e0 <+16>:    bl      0x64450 <runtime.morestack_noctxt>
   0x00000000000111e4 <+20>:    b       0x111d0 <main.(*Super).ConnectionEndpoint>
   0x00000000000111e8 <+24>:    mflr    r31
   0x00000000000111ec <+28>:    stdu    r31,-8(r1)
   0x00000000000111f0 <+32>:    addi    r3,r1,24
   0x00000000000111f4 <+36>:    addi    r3,r3,-8
   0x00000000000111f8 <+40>:    bl      0x6705c <runtime.duffzero+476>
   0x00000000000111fc <+44>:    addi    r3,r1,24
   0x0000000000011200 <+48>:    addi    r3,r3,-8
   0x0000000000011204 <+52>:    bl      0x6705c <runtime.duffzero+476>
   0x0000000000011208 <+56>:    ld      r2,16(r1)
   0x000000000001120c <+60>:    std     r2,16(r1)
   0x0000000000011210 <+64>:    ld      r2,16(r1)
   0x0000000000011214 <+68>:    cmpd    r2,r0
   0x0000000000011218 <+72>:    bne     0x11220 <main.(*Super).ConnectionEndpoint+80>
=> 0x000000000001121c <+76>:    std     r0,0(0)
   0x0000000000011220 <+80>:    b       0x11000 <main.(*Base).ConnectionEndpoint>
   0x0000000000011224 <+84>:    .long 0x0
End of assembler dump.
(gdb) bt
#0  0x000000000001121c in main.(*Super).ConnectionEndpoint (this=0x0, x=...) at <autogenerated>:2
#1  0x0000000000011128 in main.main () at /home/ubuntu/src/x/x.go:27

For some reason, when duffzero is required, it's being invoked twice.

[davecheney]

Sorry. I think I was wrong about DUFFZERO. After looking at arm64 and amd64 output, this appears to be consistent, all archs' have this double duffzero pattern. However when duffzero is not present, no corruption appears.

[mwhudson]

I see a problem in that: the tail call to the underlying method assumes that lr still points to the caller of the wrapper method, but the bl to duffzero stomps on that.

[gopherbot]

CL https://golang.org/cl/13570 mentions this issue.

[gopherbot]

CL https://golang.org/cl/14937 mentions this issue.

[gopherbot]

CL https://golang.org/cl/32181 mentions this issue.