1
Fork 0

Rollup merge of #127168 - DianQK:cast-size, r=workingjubilee

Use the aligned size for alloca at args/ret when the pass mode is cast

Fixes #75839. Fixes #121028.

The `load` and `store` instructions in LLVM access the aligned size. For example, `load { i64, i32 }` accesses 16 bytes on x86_64: https://alive2.llvm.org/ce/z/n8CHAp.

BTW, this example is expected to be optimized to immediate UB by Alive2: https://rust.godbolt.org/z/b7xK7hv1c and https://alive2.llvm.org/ce/z/vZDtZH.

r? compiler
This commit is contained in:
Matthias Krüger 2024-07-02 17:47:48 +02:00 committed by GitHub
commit 33b0238586
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
7 changed files with 381 additions and 39 deletions

View file

@ -226,7 +226,8 @@ impl<'ll, 'tcx> ArgAbiExt<'ll, 'tcx> for ArgAbi<'tcx, Ty<'tcx>> {
// when passed by value, making it smaller.
// - On some ABIs, the Rust layout { u16, u16, u16 } may be padded up to 8 bytes
// when passed by value, making it larger.
let copy_bytes = cmp::min(scratch_size.bytes(), self.layout.size.bytes());
let copy_bytes =
cmp::min(cast.unaligned_size(bx).bytes(), self.layout.size.bytes());
// Allocate some scratch space...
let llscratch = bx.alloca(scratch_size, scratch_align);
bx.lifetime_start(llscratch, scratch_size);