search

riscv-non-isa / riscv-elf-psabi-doc

A RISC-V ELF psABI Document
https://jira.riscv.org/browse/RVG-4
Creative Commons Attribution 4.0 International
699 stars 163 forks source link

alignment of arguments passed on the stack #120

Open johabu opened 5 years ago

johabu commented 5 years ago

In the integer calling convention the stack alignment is specified as follows:

Aggregates or scalars passed on the stack are aligned to the minimum of the object alignment and the stack alignment.

I suppose that arguments should be aligned to the maximum of the object alignment and the stack alignment.

Am I wrong or is this an error in the calling convention specification?

aswaterman commented 4 years ago

No, it’s correct as-is. If it were “maximum” we would have to use 16-byte stack slots for every 1-byte char passed on the stack.

On Fri, Oct 25, 2019 at 2:09 AM Johannes Bucher notifications@github.com wrote:

In the integer calling convention the stack alignment is specified as follows:

Aggregates or scalars passed on the stack are aligned to the minimum of the object alignment and the stack alignment.

I suppose that arguments should be aligned to the maximum of the object alignment and the stack alignment.

Am I wrong or is this an error in the calling convention specification?

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/riscv/riscv-elf-psabi-doc/issues/120?email_source=notifications&email_token=AAH3XQSEUVEJKAA6HWBH4W3QQKZUFA5CNFSM4JFAVE52YY3PNVWWK3TUL52HS4DFUVEXG43VMWVGG33NNVSW45C7NFSM4HUKR2ZQ, or unsubscribe https://github.com/notifications/unsubscribe-auth/AAH3XQTPETF5DRYGZN55DS3QQKZUFANCNFSM4JFAVE5Q .

johabu commented 4 years ago

I falsely assumed that "stack alignment" in this case refers to parameter stack slots of size XLEN, not the 16-byte stack alignment.

However, using "minimum" implies that the alignment for 1-byte char arguments passed on the stack is 1 byte. Multiple arguments of type char would then be placed at 0(sp), 1(sp), 2(sp), etc. Is this the intended behavior? A quick look at assembly code emitted by gcc shows that at least gcc puts them at 0(sp), 4(sp), 8(sp), etc., indicating a 4-byte alignment.

aswaterman commented 4 years ago

I forgot that the stack slots are rounded up to the word size. What I meant to say is that they aren't rounded up to the 16-byte stack alignment. So it's probably more accurate to say min(stack alignment, max(type alignment, XLEN)) or something like that.

johabu commented 4 years ago

Ok, thanks for the clarification. This is how I initially understood the calling convention, but after reading the specification again I wasn't really sure anymore because the stack slots being rounded up to the word size is not explicitly mentioned. Maybe that paragraph should be reworded to be clear.