[IR] Make @llvm.memset prototype byte width dependent

[s-barannikov]

This patch changes the type of the value argument of @llvm.memset and
similar intrinsics from i8 to iN, where N is the byte width specified
in data layout string.
Note that the argument still has fixed type (not overloaded), but type
checker will complain if the type does not match the byte width.

Ideally, the type of the argument would be dependent on the address
space of the pointer argument. It is easy to do this (and I did it
downstream as a PoC), but since data layout string doesn't currently
allow different byte widths for different address spaces, I refrained
from doing it now.

[s-barannikov]

Warning

This pull request is not mergeable via GitHub because a downstack PR is open. Once all requirements are satisfied, merge this PR as a stack on Graphite.
Learn more

• [SimplifyLibCalls] Add initial support for non-8-bit bytes #106542
• [ValueTracking] Add CharWidth argument to getConstantStringInfo (NFC) #106541
• [IR] Account for byte width in m_PtrAdd #106540
• [IRBuilder] Add getByteTy and use it in CreatePtrAdd #106539
• [ValueTracking] Make isBytewiseValue byte width agnostic #106538
• [IR] Make @llvm.memset prototype byte width dependent #106537 👈 (View in Graphite)
• [DataLayout] Add byte specification #106536
• main

This stack of pull requests is managed by Graphite. Learn more about stacking.

[nikic]

I don't think this is the right direction for memset. I would prefer actually overloading it and making the memset/memset.pattern distinction redundant. See also #97583 (comment) and following comments.

[s-barannikov]

I don't think this is the right direction for memset. I would prefer actually overloading it and making the memset/memset.pattern distinction redundant. See also #97583 (comment) and following comments.

Thanks for the link. I'm in favor of the RFC and in fact I considered this approach, too. It will take long time for the llvm.memset to be replaced with llvm.memset.pattern though, and until then we need to get llvm.memset working with non-8-bit bytes somehow. The suggested approach requires minimum changes, is backward compatible, and follows the semantics specified by the LangRef:

The ‘llvm.memset.*’ intrinsics fill “len” bytes of memory starting at the destination location.

That said, I can keep this patch downstream if you think it does something wrong.

[github-actions]

✅ With the latest revision this PR passed the C/C++ code formatter.