cmd/compile: BCE of fixed width blocks during iteration

[Jorropo]

With SIMD it is normal to iterate a slice let's say 8 elements at a time.
Sadly this incurs a bound check:

	const step = 8
	for i := 0; i <= len(x)-step; i += step {
		_ = x[i : i+step] // Unnecessary bound check here :'(
	}

[TapirLiu]

also for

	for i := 0; i <= len(x)-step; i += step {
		_ = x[i+step-1] // Unnecessary bound check here :'(
	}

see #40987

[jvoisin]

Another example, used in bcrypt's implementation:

var myArray [24]byte
for i := 0; i < 24; i += 8 {
  _ = myArray[i:i+8]  // bound check
}

[randall77]

@jvoisin Your example seems to be fixed at tip (but not 1.25rc2).
The OP's example is not fixed yet.

[slon]

I investigated the logic behind the bound check elimination, in hopes that maybe I can come up with a fix.

My understanding is that compiler tracks two kind properties:

1. Numerical range for each value, e.g. i in [0, 24].
2. Simple relations between two values, e.g a < b or a <= b.

Example from bcrypt is fixed, since it is using constant sizes. And checks are eliminated based on (1).

In the OP's example compiler would have to track relations in the form of a + K < b. So the issue looks unfixable without significant changes to (2).

[jordireinsma]

Hello, seems like this is the most active issue regarding BCEs

I have this uncovered case too:

func bce(n int) int {
	table := [5]int{4, 3, 2, 1, 0}
	i := uint32(n) % 5
	return 1 + table[i]  // bound check
}

But this one is fixed:

func bce(n int) int {
	table := [5]int{4, 3, 2, 1, 0}
	return 1 + table[uint32(n) % 5]  // no bound check
}

[Jorropo]

@jordireinsma see #63110 ✨