What is the motivation for pre/main/post loops?

[Cao, Joshua]

Had some discussion with @rkennke. A pre-loop could help for alignment.

For example, lets say we have
```
fn copy(byte[] src, byte[] dst)
  for (i = 1; i < src.length; ++i)
    dst[i] = src[i]
```

Lets say we want to align vector load/stores to words of 8 bytes. We can
assume that src/dst are word aligned. Since we start at the 1'th index, we
know the first few iterations are not word aligned and can put them in the
beginning loop. We can transform it into:

```
// pre-loop
for (i = 1; i < 8; i++) { dst[i] = src[i] }
// main-loop
for (i = 8; i < src.length - (src.length % 16); i+=16) {
  // 128 bit / 16 byte / 2 word vector load-store
  dst[i:i+16] = src[i:i+16]
}
// post-loop, remainder iterations
for (i; i < src.length; i++) { dst[i] = src[i] }
```

I found that clang also generates a pre-loop for the same example. They call it
“prolog/main/epilog” loop https://godbolt.org/z/KrEfj7z56.

I would like to hear if there are any other reasons for a pre-loop.

________________________________
From: Cao, Joshua
Sent: Wednesday, February 7, 2024 10:53:09 AM
To: hotspot-compiler-dev at openjdk.org
Subject: What is the motivation for pre/main/post loops?


I can't find any notes on this. The closest I can find is https://wiki.openjdk.org/pages/viewpage.action?pageId=20415918


> clone loop body and create pre-, main- and post- loops


I can guess why a post loop is useful. We can use it for the remainder iterations after loop unrolling.

However, I cannot think of scenarios where a pre-loop is useful. What are some cases where a

pre-loop is useful?
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://mail.openjdk.org/pipermail/hotspot-compiler-dev/attachments/20240208/303fd37f/attachment-0001.htm>