rustc_const_eval: respect target.min_global_align

[folkertdev]

This PR builds on #142179 (which will be merged soon), so ignore the first commit here.

Issue #44411 describes a correctness problem on s390x, where a static is not sufficiently aligned: the larl instruction that LLVM generates assumes an even address, but e.g. a static containing a bool may have an odd address.

Because miri is an interpreter, it is not actually behaving incorrectly in this case. Currently, the only LLVM target that also specifies a min_global_align is larai, which rust does not support. The nvptx targets inherit the alignment from their host, but nvptx + s390x seems unlikely. Nevertheless, miri should probably respect the target's min_global_align.

The solution here is really hard to test: it's kind of a numbers game. The test I added at least fails on current miri (and should now succeed on CI). We'll likely soon have #[align] on statics (rust-lang/rfcs#3806 is in FCP), so then we can have some more extensive static alignment checks. I'm assuming the tests are executed for a variety of targets, including s390x?

See also #miri > alignment of globals on s390x

r? @RalfJung

[rustbot]

The Miri subtree was changed

cc @rust-lang/miri

These commits modify compiler targets.
(See the Target Tier Policy.)

Some changes occurred to the CTFE machinery

cc @RalfJung, @oli-obk, @lcnr

Some changes occurred in compiler/rustc_codegen_gcc

cc @antoyo, @GuillaumeGomez

Some changes occurred to the CTFE / Miri interpreter

cc @rust-lang/miri

[saethlin]

"in practice" seems like unnecessarily wobbly language here. Can we say something more concrete, or if the answer is complicated point to where the full answer is?

[workingjubilee]

the actual answer is "LLVM and some other compilers generate incorrect code if globals have an alignment less than 2, as they generate accesses to globals using LALR, which requires alignment to even addresses in order to work".

[folkertdev]

Yeah, the situation here is a confusing

• technically, miri does not need to enforce this alignment of 2, because it does not actually emit the larl instruction that would behave incorrectly
• but in practice, both gcc, clang and rustc do respect this alignment
• intuitively miri should respect target.min_global_align
• the only way to observe its effect right now is to test the alignment of statics on s390x

Anyway I've updated the comment with why this test was written.