Improve pointer arithmetic docs
* Add slightly more detailed definition of "allocated object" to the module docs, and link it from everywhere.
* Clarify the "remains attached" wording a bit (at least I hope this is clearer).
* Remove the sentence about using integer arithmetic; this seems to confuse people even if it is technically correct.
As usual, the edit needs to be done in a dozen places to remain consistent, I hope I got them all.
The use of module-level functions instead of associated functions
on `<*const T>` or `<*mut T>` follows the precedent of
`ptr::slice_from_raw_parts` and `ptr::slice_from_raw_parts_mut`.
clarify rules for ZST Boxes
LLVM's rules around `getelementptr inbounds` with offset 0 are a bit annoying, and as a consequence we have no choice but say that a `Box<()>` pointing to previously allocated memory that has since been freed is UB. Clarify the docs to reflect this.
This is based on conversations on the LLVM mailing list.
* Here's my initial mail: https://lists.llvm.org/pipermail/llvm-dev/2019-February/130452.html
* The first email of the March part of that thread: https://lists.llvm.org/pipermail/llvm-dev/2019-March/130831.html
* First email of the April part: https://lists.llvm.org/pipermail/llvm-dev/2019-April/131693.html
The conclusion for me at least was that `getelementptr inbounds` with offset 0 is *not* the identity function, but can sometimes return `poison` even when the input is a regular pointer -- specifically, it returns `poison` when this pointer points into something that LLVM "knows has been deallocated", i.e., a former LLVM-managed allocation. It is however the identity function on pointers obtained by casting integers.
Note that there [are formal proposals](https://people.mpi-sws.org/~jung/twinsem/twinsem.pdf) for LLVM semantics where `getelementptr inbounds` with offset 0 isn't quite the identity function but never returns `poison` (it affects the provenance of the pointer but in a way that doesn't matter if this pointer is never used for memory accesses), and indeed this is likely necessary to consistently describe LLVM semantics. But with the informal LLVM LangRef that we have right now, and with LLVM devs insisting otherwise, it seems unwise to rely on this.
Optimise align_offset for stride=1 further
`stride == 1` case can be computed more efficiently through `-p (mod
a)`. That, then translates to a nice and short sequence of LLVM
instructions:
%address = ptrtoint i8* %p to i64
%negptr = sub i64 0, %address
%offset = and i64 %negptr, %a_minus_one
And produces pretty much ideal code-gen when this function is used in
isolation.
Typical use of this function will, however, involve use of
the result to offset a pointer, i.e.
%aligned = getelementptr inbounds i8, i8* %p, i64 %offset
This still looks very good, but LLVM does not really translate that to
what would be considered ideal machine code (on any target). For example
that's the codegen we obtain for an unknown alignment:
; x86_64
dec rsi
mov rax, rdi
neg rax
and rax, rsi
add rax, rdi
In particular negating a pointer is not something that’s going to be
optimised for in the design of CISC architectures like x86_64. They
are much better at offsetting pointers. And so we’d love to utilize this
ability and produce code that's more like this:
; x86_64
lea rax, [rsi + rdi - 1]
neg rsi
and rax, rsi
To achieve this we need to give LLVM an opportunity to apply its
various peep-hole optimisations that it does during DAG selection. In
particular, the `and` instruction appears to be a major inhibitor here.
We cannot, sadly, get rid of this load-bearing operation, but we can
reorder operations such that LLVM has more to work with around this
instruction.
One such ordering is proposed in #75579 and results in LLVM IR that
looks broadly like this:
; using add enables `lea` and similar CISCisms
%offset_ptr = add i64 %address, %a_minus_one
%mask = sub i64 0, %a
%masked = and i64 %offset_ptr, %mask
; can be folded with `gepi` that may follow
%offset = sub i64 %masked, %address
…and generates the intended x86_64 machine code.
One might also wonder how the increased amount of code would impact a
RISC target. Turns out not much:
; aarch64 previous ; aarch64 new
sub x8, x1, #1 add x8, x1, x0
neg x9, x0 sub x8, x8, #1
and x8, x9, x8 neg x9, x1
add x0, x0, x8 and x0, x8, x9
(and similarly for ppc, sparc, mips, riscv, etc)
The only target that seems to do worse is… wasm32.
Onto actual measurements – the best way to evaluate snipets like these
is to use llvm-mca. Much like Aarch64 assembly would allow to suspect,
there isn’t any performance difference to be found. Both snippets
execute in same number of cycles for the CPUs I tried. On x86_64,
we get throughput improvement of >50%!
Fixes#75579
`write` is ambiguous because there's also a macro called `write`.
Also removed unnecessary and potentially confusing link to a function in
its own docs.
The only link that I did not change is a link to a function on the
`pointer` primitive because intra-doc links for the `pointer` primitive
don't work yet (see #63351).
`stride == 1` case can be computed more efficiently through `-p (mod
a)`. That, then translates to a nice and short sequence of LLVM
instructions:
%address = ptrtoint i8* %p to i64
%negptr = sub i64 0, %address
%offset = and i64 %negptr, %a_minus_one
And produces pretty much ideal code-gen when this function is used in
isolation.
Typical use of this function will, however, involve use of
the result to offset a pointer, i.e.
%aligned = getelementptr inbounds i8, i8* %p, i64 %offset
This still looks very good, but LLVM does not really translate that to
what would be considered ideal machine code (on any target). For example
that's the codegen we obtain for an unknown alignment:
; x86_64
dec rsi
mov rax, rdi
neg rax
and rax, rsi
add rax, rdi
In particular negating a pointer is not something that’s going to be
optimised for in the design of CISC architectures like x86_64. They
are much better at offsetting pointers. And so we’d love to utilize this
ability and produce code that's more like this:
; x86_64
lea rax, [rsi + rdi - 1]
neg rsi
and rax, rsi
To achieve this we need to give LLVM an opportunity to apply its
various peep-hole optimisations that it does during DAG selection. In
particular, the `and` instruction appears to be a major inhibitor here.
We cannot, sadly, get rid of this load-bearing operation, but we can
reorder operations such that LLVM has more to work with around this
instruction.
One such ordering is proposed in #75579 and results in LLVM IR that
looks broadly like this:
; using add enables `lea` and similar CISCisms
%offset_ptr = add i64 %address, %a_minus_one
%mask = sub i64 0, %a
%masked = and i64 %offset_ptr, %mask
; can be folded with `gepi` that may follow
%offset = sub i64 %masked, %address
…and generates the intended x86_64 machine code. One might also wonder
how the increased amount of code would impact a RISC target. Turns out
not much:
; aarch64 previous ; aarch64 new
sub x8, x1, #1 add x8, x1, x0
neg x9, x0 sub x8, x8, #1
and x8, x9, x8 neg x9, x1
add x0, x0, x8 and x0, x8, x9
(and similarly for ppc, sparc, mips, riscv, etc)
The only target that seems to do worse is… wasm32.
Onto actual measurements – the best way to evaluate snippets like these
is to use llvm-mca. Much like Aarch64 assembly would allow to suspect,
there isn’t any performance difference to be found. Both snippets
execute in same number of cycles for the CPUs I tried. On x86_64,
we get throughput improvement of >50%, however!
Previously checking for `pmoda == 0` would get LLVM to generate branchy
code, when, for `stride = 1` the offset can be computed without such a
branch by doing effectively a `-p % a`.
For well-known (constant) alignments, with the new ordering of these
conditionals, we end up generating 2 to 3 cheap instructions on x86_64:
movq %rdi, %rax
negl %eax
andl $7, %eax
instead of 5+ as previously.
For unknown alignments the new code also generates just 3 instructions:
negq %rdi
leaq -1(%rsi), %rax
andq %rdi, %rax
At opt-level <= 1, the methods such as `wrapping_mul` are not being
inlined, causing significant bloating and slowdowns of the
implementation at these optimisation levels.
With use of these intrinsics, the codegen of this function at
-Copt_level=1 is the same as it is at -Copt_level=3.
