[llvm] [libsycl] USM Aligned allocation functions (PR #213468)

[via llvm-commits]

================
@@ -103,14 +154,29 @@ void *malloc(std::size_t numBytes, const device &syclDevice,
 
   void *Ptr{};
   auto OLDevice = detail::getSyclObjImpl(syclDevice)->getOLHandle();
-  auto Result =
-      kind == usm::alloc::host
-          ? detail::callNoCheck(olMemAllocHost, OLDevice, numBytes, &Ptr)
-          : detail::callNoCheck(olMemAlloc, OLDevice,
-                                detail::getOlAllocType(kind), numBytes, &Ptr);
+  auto Result = kind == usm::alloc::host
+                    ? detail::callNoCheck(olMemAllocAlignedHost, OLDevice,
----------------
Robertkq wrote:

Ok.. so I have a follow up just so it's clearer for me in which way to take development

Throughout the code, as you also noted, I have certain `modifications` of alignment argument, that tries to address [Alignment guarantees of USM allocation functions ](https://registry.khronos.org/SYCL/specs/sycl-2020/html/sycl-2020.html#_usm_allocations)

I believe that if we were to fully support this, which I don't necessarily see why not, we could use only the `aligned` versions of `olMemAlloc` for USM functions, because all types of allocations need guarantees which can be explicitly satisfied with `olMemAllocAligned` rather than implicitly with `olMemAlloc`

If possible, can you suggest a path I should go ahead for the implementation with?

I could rather:

1. Remove explicit alignment guarantees & all memory allocations requested with 0 alignment will call `olMemAlloc` 
> This option make use of more `offload` API which could be "clearer", but we implicitly rely on liboffload for alignment guarantees 

2. Add necessary guarantees to also satisfy row 3 of Table n. 70, keep routing everything via `olMemAllocAligned`, `aligned_alloc` will always have alignment ` != 0` by the time it reaches it's last overload, even if user called with `0`, it will change to minimum guarantee
> I believe this option gives us explicit guarantees to conform to SYCL specs

Hope I understood well and the question is well formed

https://github.com/llvm/llvm-project/pull/213468