cmd/compile: optimized range memclr doesn't work with pointers to arrays

[josharian]

package p

type T struct {
	a *[10]int
	b [10]int
}

func (t *T) resetA() {
	for i := range t.a {
		t.a[i] = 0
	}
}

func (t *T) resetB() {
	for i := range t.b {
		t.b[i] = 0
	}
}

resetA compiles to a loop. resetB compiles to a call to runtime.memclrNoHeapPointers. I believe that resetA should also call memclrNoHeapPointers.

This showed up as a hot spot in my code today (with much larger arrays).

cc @randall77 @cuonglm @mdempsky

[josharian]

As evidence that this is just a missed optimization, here is equivalent code using a new intermediate type:

type T struct {
	c *A
}

type A struct {
	x [10]int
}

func (t *T) resetC() {
	for i := range t.c.x {
		t.c.x[i] = 0
	}
}

This generates runtime.memclrNoHeapPointers. (And also provides a workaround.)

[renthraysk]

There is a simpler workaround for i := range *t.a {.

[josharian]

@renthraysk I am astonished that that works, and I wrote the original optimization. Now that you say it, I can imagine why, although it has to do with weird quirks in the guts of the compiler. Anyway, I still think this is worth fixing. :P

[renthraysk]

Yea, agreed. Actually thought I filed an issue for this, but seems just tweeted about it back in 2017, so my bad.

Subtle diff in
func f1(in *[96]byte) { for i := range in { in[i] = 0 } }
func f2(in *[96]byte) { for i := range *in { in[i] = 0 } }

[gopherbot]

Change https://go.dev/cl/403337 mentions this issue: cmd/compile: support pointers to arrays in arrayClear

[rip-create-your-account]

The following generates slightly more efficient code for my computer.

func (t *T) resetC() {
	*t.a = [10]int{}
}

There's no runtime.memclrNoHeapPointers call at all. Instead the compiler seems to choose a suitable algorithm based on the size of the array. To me it seems like all of the loops above should compile to this.

[zigo101]

It looks the memclr way is more efficient than the assignment way for small arrays.
Should the runtime internal unify the implementation for the two ways?