I understand that it does so when the exact type is observed - a direct call on a concrete type. But I was wondering if it performs whole-program-view optimization for interface calls. E.g. given a simple AOT-compiled C# program:
using System.Runtime.CompilerServices;
var bar = new Bar();
var number = CallFoo(bar);
Console.WriteLine(number);
// Do not inline to prevent observing exact type
[MethodImpl(MethodImplOptions.NoInlining)]
static int CallFoo(Foo foo) {
return foo.Number();
}
interface Foo {
int Number();
}
class Bar: Foo {
public int Number() => 42;
}
On x86_64, 'CallFoo' compiles to:
CMP byte ptr [RDI],DIL ;; null-check foo[0]
MOV EAX,0x2a ;; set 42 to return value register
RET
There is no interface call. In the above case, the linker reasons that throughout whole program only `Bar` implements `Foo` therefore all calls on `Foo` can be replaced with direct calls on `Bar`, which are then subject to other optimizations like inlining.
In fact, if we add and reference a second implementation of `Foo` - `Baz` which returns 8, `CallFoo` becomes
;; calculate the addr. of Bar's methodtable pointer
LEA RAX,[DevirtExample_Bar::vtable]
MOV ECX,0x8 ;; set ECX to 8
MOV EDX,0x2a ;; set EDX to 42
;; compare methodtable pointer of foo instance with Bar's
CMP qword ptr [RDI],RAX
;; set return register EAX to value of EDX, containing 42
MOV EAX,EDX
;; if comparison is false, set EAX to value of ECX containing 8 instead
CMOVNZ EAX,ECX
RET
Which is effectively 'return foo is Bar ? 42 : 8;'.
Despite my criticism of Go's capabilities, I am interested in how its implementation is evolving. I know it has the feature to manually gather a static PGO profile and then apply it to compilation which will insert guarded devirtualization fast-paths on interface calls, like what OpenJDK's HotSpot and .NET's JIT do automatically. But I was wondering whether it was doing any whole-program view or inter-procedural optimizations that can be very effective with "frozen world single static module" which both Go and .NET AOT compilations are.
EDIT: To answer my own question, I verified the same for Go. Given simple Go program:
package main
import (
"fmt"
)
func main() {
bar := &Bar{}
num1 := callFoo(bar)
fmt.Println(num1)
}
//go:noinline
func callFoo(foo Foo) int {
return foo.Number()
}
type Foo interface {
Number() int
}
type Bar struct{}
func (b *Bar) Number() int {
return 42
}
'callFoo' compiles to
CMP RSP,qword ptr [R14 + 0x10]
JBE LAB_0108ca68
PUSH RBP
MOV RBP,RSP
SUB RSP,0x8
MOV qword ptr [RSP + foo_spill.tab],RAX
MOV qword ptr [RSP + foo_spill.data],RBX
MOV RCX,qword ptr [RAX + 0x18] ;; load vtable slot?
MOV RAX,RBX
NOP
CALL RCX ;; call the address loaded from the vtable?
ADD RSP,0x8
POP RBP
RET
LAB_0108ca68 XREF[1]:
MOV qword ptr [RSP + foo_spill.tab],RAX
MOV qword ptr [RSP + foo_spill.data],RBX
CALL runtime.morestack_noctxt
MOV RAX,qword ptr [RSP + foo_spill.tab]
MOV RBX,qword ptr [RSP + foo_spill.data]
JMP main.callFoo
It appears that no devirtualization takes place of this kind. Writing about this, it makes for an interesting thought experiment what it would take to introduce a CIL back-end for Go (including proper export of types, and what about structurally matched interfaces?) and AOT compile it with .NET.
[0]: VMs like OpenJDK and .NET make hardware exception-based null-checks. That is, a SIGSEGV handler is registered and then pointers that need to throw NRE or NPE either do so via induced loads from memory like above or just by virtue of dereferencing a field out of an object reference. If a pointer is null, this causes SIGSEGV, where then a handler looks if the address of the invalid pointer is within first, say, 64KiB of address space. If it is, the VM logic kicks in that recovers the execution state and performs managed exception handling such as running `finally` blocks and resuming the execution from the corresponding `catch` handler.