Make strict provenance compatible #542
Conversation
src/bytes.rs
Outdated
| fn ptr_map<F>(ptr: *mut u8, f: F) -> *mut u8 | ||
| where | ||
| F: FnOnce(usize) -> usize, | ||
| { | ||
| let old_addr = ptr as usize; | ||
| let new_addr = f(old_addr); | ||
| let diff = new_addr.wrapping_sub(old_addr); | ||
| ptr.wrapping_add(diff) | ||
| } |
There was a problem hiding this comment.
Unfortunately, it seems like this doesn't really optimize as well as I had hoped. You can view godbolt here to see the difference.
There was a problem hiding this comment.
One option is to go through a normal pointer cast for non-miri release mode.
There was a problem hiding this comment.
Is this valid? This seems to optimize well, though I have no idea why. I just tried it out because I find this pattern for writing pointer packing more intuitive.
https://rust.godbolt.org/z/a949Wjs7f (I previously left a target-cpu in the flags, oops)
|
FYI, the optimization you do to avoid the bad codegen causes (non-theoretical) LLVM miscompiles -- See rust-lang/rust#96538 for some details. So I don't know if you should keep it. |
|
welp. that's cool. |
|
@thomcc Thanks. We will uses ptr2int2ptr casts for now on non-miri builds then. |
This PR replaces all ptr2int2ptr casts with an
wrapping_addequivalent, making it strict-provenance compatible.