Cranelift: add alignment parameter to stack slots. #8635
Conversation
Fixes bytecodealliance#6716. Currently, stack slots on the stack are aligned only to a machine-word boundary. This is insufficient for some use-cases: for example, storing SIMD data or structs that require a larger alignment. This PR adds a parameter to the `StackSlotData` to specify alignment, and the associated logic to the CLIF parser and printer. It updates the shared ABI code to compute the stackslot layout taking the alignment into account. In order to ensure the alignment is always a power of two, it is stored as a shift amount (log2 of actual alignment) in the IR.
cfallin
force-pushed
the
stackslot-alignment
branch
from
0855496 to
6d96151
Compare
|
(for the record: work on this now spawned by this thread) |
Co-authored-by: Trevor Elliott <awesomelyawesome@gmail.com>
| // satisfy the user's requested alignment. | ||
| debug_assert!(data.align_shift < 32); | ||
| let align = std::cmp::max(M::word_bytes(), 1u32 << data.align_shift); | ||
| let mask = align - 1; |
There was a problem hiding this comment.
If the alignment is bigger than the ABI stack alignment it would be necessary to adjust the stack pointer to correctly align it. This is missing here. Rust allows arbitrary power of two alignments for stack variables, even something crazy like a 1GB alignment. More realistically though, cache line alignment is used by crates like rayon for performance reasons. Without correctly aligning, rustc will trigger a debug assertion in some cases like for the aforementioned rayon crate.
There was a problem hiding this comment.
Ah, yes, that's a good point -- to summarize: alignment of offset within the stack frame to alignment A does not imply alignment to A overall, because SP itself may be aligned less (e.g. only to A/4).
I don't have time to work further on this at the moment, but I'd be happy to review a PR that addressed this.
Fixes #6716.
Currently, stack slots on the stack are aligned only to a machine-word boundary. This is insufficient for some use-cases: for example, storing SIMD data or structs that require a larger alignment.
This PR adds a parameter to the
StackSlotDatato specify alignment, and the associated logic to the CLIF parser and printer. It updates the shared ABI code to compute the stackslot layout taking the alignment into account. In order to ensure the alignment is always a power of two, it is stored as a shift amount (log2 of actual alignment) in the IR.