IR: Taking function arg address and representing references as pointers #6967
Conversation
…to vaivaswatha/ir_consts_1
…to vaivaswatha/ir_consts_2
…to vaivaswatha/ir_consts_2
This is the third PR to address #6351. It allows safely taking addresses of function arguments.
vaivaswatha
temporarily deployed
to
fuel-sway-bot
GitHub Actions
Inactive
vaivaswatha
force-pushed
the
vaivaswatha/ir2_pr3
branch
from
9bb2b0c to
e65a8da
Compare
vaivaswatha
temporarily deployed
to
fuel-sway-bot
GitHub Actions
Inactive
vaivaswatha
temporarily deployed
to
fuel-sway-bot
GitHub Actions
Inactive
vaivaswatha
temporarily deployed
to
fuel-sway-bot
GitHub Actions
Inactive
There was a problem hiding this comment.
Great job @vaivaswatha! 💪 🚀
To my comments, those regarding replacing asm code generated for slices, can IMO be addressed in a separate PR.
Looking forward to get this PR merged and to continue the work on references!
ironcev
changed the title
ironcev
added
compiler
Thank you @IGI-111 . This was a tough one !
Yes, I agree with you about the comments, and that it's better done as a subsequent PR. Those ASM blocks were sources of multiple bugs that took me quite a while to get right, each time with me wondering why we have it like that.
Indeed ! I'll work on the other comments that you have and update the PR soon. From a first look, I agree with them all. Thank you. |
Co-authored-by: Igor Rončević <[email protected]>
vaivaswatha
temporarily deployed
to
fuel-sway-bot
GitHub Actions
Inactive
There was a problem hiding this comment.
One more thought, in the verify_ptr_to_int, I've enabled the check to make sure that we don't have any more semantically invalid PtrToInt casts, where the pointer argument is actually not the pointer.
Good news - looks like it is so 😄
"Bad" news - there are failing tests, but they all fail in the same case, on casting string slices to int. But this is more a conceptual thing because string<N> in IR is actually a pointer to the slice.
This is IMO a subject for an upcoming PR that will extend the verifyer.
Here is an example of such issues in tests, they are all essentially the same:
v6 = const string<4> "main"
v7 = ptr_to_int v6 to u64
The failing conversion will in this case be:
Internal compiler error: Verification failed: Pointer cast from non pointer string<4>.
Description
This is the third PR to address #6351. It allows safely taking addresses of function arguments, and represents references (
&) as pointers in the IR.Note: The entire function
type_correction(inir_generation.rs) and this edit incompile.rs, to handle return values, must be removed once we fix and finalize the syntax of references. Until then these two hacks need to stay. Ideally I wanted to put them both in one place (which is thetype_correctionpass), but the return type fix needs to happen, unfortunately, during IR gen.