Vector operations can produce underread on allocated buffers

[steven-johnson]

Some failures in #3252 appear to be happening because we can generate code that will do vector loads from before the allocation start of a malloc'ed buffer; in the past, this has never been an issue because our halide_malloc implementation overallocates at the start in order to ensure proper alignment (so the 'illegal' read is still from allocated memory). The change to use an already-aligned alloc exposes this issue; it appears that the same issue is triggering at least some failures in the experimental WASM branch in conjunction with the JIT tests (which is also using an allocator with no 'slop' at the beginning).

Options:

• Fix codegen so that we can never underread an allocated buffer.
• Explicitly quantify the apron sizes needed both before and after an allocated buffer, and ensure that all allocators follow the rules (either via better code in the caller or by stricter implementation requirements in halide_malloc).

[steven-johnson]

Did this really get fixed? I think it's still an issue.

[dsharletg]

It looks like it probably is. I saw the reference from #1044 last, but it looks like that was also a stale issue being closed, not the PR that I thought it was.