Add `select_unpredictable` to force LLVM to use CMOV
Since https://reviews.llvm.org/D118118, LLVM will no longer turn CMOVs into branches if it comes from a `select` marked with an `unpredictable` metadata attribute.
This PR introduces `core::intrinsics::select_unpredictable` which emits such a `select` and uses it in the implementation of `binary_search_by`.
Since https://reviews.llvm.org/D118118, LLVM will no longer turn CMOVs
into branches if it comes from a `select` marked with an `unpredictable`
metadata attribute.
This PR introduces `core::intrinsics::select_unpredictable` which emits
such a `select` and uses it in the implementation of `binary_search_by`.
Except for `simd-intrinsic/`, which has a lot of files containing
multiple types like `u8x64` which really are better when hand-formatted.
There is a surprising amount of two-space indenting in this directory.
Non-trivial changes:
- `rustfmt::skip` needed in `debug-column.rs` to preserve meaning of the
test.
- `rustfmt::skip` used in a few places where hand-formatting read more
nicely: `enum/enum-match.rs`
- Line number adjustments needed for the expected output of
`debug-column.rs` and `coroutine-debug.rs`.
Stop using LLVM struct types for alloca
The alloca type has no semantic meaning, only the size (and alignment, but we specify it explicitly) matter. Using `[N x i8]` is a more direct way to specify that we want `N` bytes, and avoids relying on LLVM's struct layout. It is likely that a future LLVM version will change to an untyped alloca representation.
Split out from #121577.
r? `@ghost`
Dellvmize some intrinsics (use `u32` instead of `Self` in some integer intrinsics)
This implements https://github.com/rust-lang/compiler-team/issues/693 minus what was implemented in #123226.
Note: I decided to _not_ change `shl`/... builder methods, as it just doesn't seem worth it.
r? ``@scottmcm``
Add `Ord::cmp` for primitives as a `BinOp` in MIR
Update: most of this OP was written months ago. See https://github.com/rust-lang/rust/pull/118310#issuecomment-2016940014 below for where we got to recently that made it ready for review.
---
There are dozens of reasonable ways to implement `Ord::cmp` for integers using comparison, bit-ops, and branches. Those differences are irrelevant at the rust level, however, so we can make things better by adding `BinOp::Cmp` at the MIR level:
1. Exactly how to implement it is left up to the backends, so LLVM can use whatever pattern its optimizer best recognizes and cranelift can use whichever pattern codegens the fastest.
2. By not inlining those details for every use of `cmp`, we drastically reduce the amount of MIR generated for `derive`d `PartialOrd`, while also making it more amenable to MIR-level optimizations.
Having extremely careful `if` ordering to μoptimize resource usage on broadwell (#63767) is great, but it really feels to me like libcore is the wrong place to put that logic. Similarly, using subtraction [tricks](https://graphics.stanford.edu/~seander/bithacks.html#CopyIntegerSign) (#105840) is arguably even nicer, but depends on the optimizer understanding it (https://github.com/llvm/llvm-project/issues/73417) to be practical. Or maybe [bitor is better than add](https://discourse.llvm.org/t/representing-in-ir/67369/2?u=scottmcm)? But maybe only on a future version that [has `or disjoint` support](https://discourse.llvm.org/t/rfc-add-or-disjoint-flag/75036?u=scottmcm)? And just because one of those forms happens to be good for LLVM, there's no guarantee that it'd be the same form that GCC or Cranelift would rather see -- especially given their very different optimizers. Not to mention that if LLVM gets a spaceship intrinsic -- [which it should](https://rust-lang.zulipchat.com/#narrow/stream/131828-t-compiler/topic/Suboptimal.20inlining.20in.20std.20function.20.60binary_search.60/near/404250586) -- we'll need at least a rustc intrinsic to be able to call it.
As for simplifying it in Rust, we now regularly inline `{integer}::partial_cmp`, but it's quite a large amount of IR. The best way to see that is with 8811efa88b (diff-d134c32d028fbe2bf835fef2df9aca9d13332dd82284ff21ee7ebf717bfa4765R113) -- I added a new pre-codegen MIR test for a simple 3-tuple struct, and this PR change it from 36 locals and 26 basic blocks down to 24 locals and 8 basic blocks. Even better, as soon as the construct-`Some`-then-match-it-in-same-BB noise is cleaned up, this'll expose the `Cmp == 0` branches clearly in MIR, so that an InstCombine (#105808) can simplify that to just a `BinOp::Eq` and thus fix some of our generated code perf issues. (Tracking that through today's `if a < b { Less } else if a == b { Equal } else { Greater }` would be *much* harder.)
---
r? `@ghost`
But first I should check that perf is ok with this
~~...and my true nemesis, tidy.~~
Making `libcore` decide this is silly; the backend has so much better information about when it's a good idea.
So introduce a new `typed_swap` intrinsic with a fallback body, but replace that implementation for immediates and scalar pairs.
Currently, we assume that ScalarPair is always represented using
a two-element struct, both as an immediate value and when stored
in memory.
This currently works fairly well, but runs into problems with
https://github.com/rust-lang/rust/pull/116672, where a ScalarPair
involving an i128 type can no longer be represented as a two-element
struct in memory. For example, the tuple `(i32, i128)` needs to be
represented in-memory as `{ i32, [3 x i32], i128 }` to satisfy
alignment requirement. Using `{ i32, i128 }` instead will result in
the second element being stored at the wrong offset (prior to
LLVM 18).
Resolve this issue by no longer requiring that the immediate and
in-memory type for ScalarPair are the same. The in-memory type
will now look the same as for normal struct types (and will include
padding filler and similar), while the immediate type stays a
simple two-element struct type. This also means that booleans in
immediate ScalarPair are now represented as i1 rather than i8,
just like we do everywhere else.
The core change here is to llvm_type (which now treats ScalarPair
as a normal struct) and immediate_llvm_type (which returns the
two-element struct that llvm_type used to produce). The rest is
fixing things up to no longer assume these are the same. In
particular, this switches places that try to get pointers to the
ScalarPair elements to use byte-geps instead of struct-geps.
These tend to have special handling in a bunch of places anyway, so the variant helps remember that. And I think it's easier to grok than non-Scalar Aggregates sometimes being `Immediates` (like I got wrong and caused 109992). As a minor bonus, it means we don't need to generate poison LLVM values for them to pass around in `OperandValue::Immediate`s.
This takes a whole 3 lines in `compiler/` since it lowers to `CastKind::Transmute` in MIR *exactly* the same as the existing `intrinsics::transmute` does, it just doesn't have the fancy checking in `hir_typeck`.
Added to enable experimenting with the request in <https://github.com/rust-lang/rust/pull/106281#issuecomment-1496648190> and because the portable-simd folks might be interested for dependently-sized array-vector conversions.
It also simplifies a couple places in `core`.
LLVM can usually optimize these away, but especially for things like transmutes of newtypes it's silly to generate the `alloc`+`store`+`load` at all when it's actually a nop at LLVM level.
Updates `interpret`, `codegen_ssa`, and `codegen_cranelift` to consume the new cast instead of the intrinsic.
Includes `CastTransmute` for custom MIR building, to be able to test the extra UB.
Previously, it was only put on scalars with range validity invariants
like bool, was uninit was obviously invalid for those.
Since then, we have normatively declared all uninit primitives to be
undefined behavior and can therefore put `noundef` on them.
The remaining concern was the `mem::uninitialized` function, which cause
quite a lot of UB in the older parts of the ecosystem. This function now
doesn't return uninit values anymore, making users of it safe from this
change.
The only real sources of UB where people could encounter uninit
primitives are `MaybeUninit::uninit().assume_init()`, which has always
be clear in the docs about being UB and from heap allocations (like
reading from the spare capacity of a vec. This is hopefully rare enough
to not break anything.
