Stop using LLVM struct types for byval/sret
For `byval` and `sret`, the type has no semantic meaning, only the size matters\*†. Using `[N x i8]` is a more direct way to specify that we want `N` bytes, and avoids relying on LLVM's struct layout.
\*: The alignment would matter, if we didn't explicitly specify it. From what I can tell, we always specified the alignment for `sret`; for `byval`, we didn't until #112157.
†: For `byval`, the hidden copy may be impacted by padding in the LLVM struct type, i.e. padding bytes may not be copied. (I'm not sure if this is done today, but I think it would be legal.) But we manually pad our LLVM struct types specifically to avoid there ever being LLVM-visible padding, so that shouldn't be an issue.
Split out from #121577.
r? `@nikic`
Currently, we assume that ScalarPair is always represented using
a two-element struct, both as an immediate value and when stored
in memory.
This currently works fairly well, but runs into problems with
https://github.com/rust-lang/rust/pull/116672, where a ScalarPair
involving an i128 type can no longer be represented as a two-element
struct in memory. For example, the tuple `(i32, i128)` needs to be
represented in-memory as `{ i32, [3 x i32], i128 }` to satisfy
alignment requirement. Using `{ i32, i128 }` instead will result in
the second element being stored at the wrong offset (prior to
LLVM 18).
Resolve this issue by no longer requiring that the immediate and
in-memory type for ScalarPair are the same. The in-memory type
will now look the same as for normal struct types (and will include
padding filler and similar), while the immediate type stays a
simple two-element struct type. This also means that booleans in
immediate ScalarPair are now represented as i1 rather than i8,
just like we do everywhere else.
The core change here is to llvm_type (which now treats ScalarPair
as a normal struct) and immediate_llvm_type (which returns the
two-element struct that llvm_type used to produce). The rest is
fixing things up to no longer assume these are the same. In
particular, this switches places that try to get pointers to the
ScalarPair elements to use byte-geps instead of struct-geps.
alignment of `byval` on x86 in the process.
Commit 88e4d2c291 from five years ago removed
support for alignment on indirectly-passed arguments because of problems with
the `i686-pc-windows-msvc` target. Unfortunately, the `memcpy` optimizations I
recently added to LLVM 16 depend on this to forward `memcpy`s. This commit
attempts to fix the problems with `byval` parameters on that target and now
correctly adds the `align` attribute.
The problem is summarized in [this comment] by @eddyb. Briefly, 32-bit x86 has
special alignment rules for `byval` parameters: for the most part, their
alignment is forced to 4. This is not well-documented anywhere but in the Clang
source. I looked at the logic in Clang `TargetInfo.cpp` and tried to replicate
it here. The relevant methods in that file are
`X86_32ABIInfo::getIndirectResult()` and
`X86_32ABIInfo::getTypeStackAlignInBytes()`. The `align` parameter attribute
for `byval` parameters in LLVM must match the platform ABI, or miscompilations
will occur. Note that this doesn't use the approach suggested by eddyb, because
I felt it was overkill to store the alignment in `on_stack` when special
handling is really only needed for 32-bit x86.
As a side effect, this should fix#80127, because it will make the `align`
parameter attribute for `byval` parameters match the platform ABI on LLVM
x86-64.
[this comment]: https://github.com/rust-lang/rust/pull/80822#issuecomment-829985417
Previously, it was only put on scalars with range validity invariants
like bool, was uninit was obviously invalid for those.
Since then, we have normatively declared all uninit primitives to be
undefined behavior and can therefore put `noundef` on them.
The remaining concern was the `mem::uninitialized` function, which cause
quite a lot of UB in the older parts of the ecosystem. This function now
doesn't return uninit values anymore, making users of it safe from this
change.
The only real sources of UB where people could encounter uninit
primitives are `MaybeUninit::uninit().assume_init()`, which has always
be clear in the docs about being UB and from heap allocations (like
reading from the spare capacity of a vec. This is hopefully rare enough
to not break anything.
