Returns values up to 2*usize by value #77434
Conversation
|
r? @estebank (rust_highfive has picked a reviewer for you, use r? to override) |
rust-highfive
added
the
S-waiting-on-review
|
r? @nagisa |
|
@bors try @rust-timer queue cc @eddyb |
|
Awaiting bors try build completion |
|
⌛ Trying commit b01694e with merge af826c3de514b426f1172de2883db4a251e45c65... |
|
Perf runs only on x86_64 so there should be no noticeable change. |
|
☀️ Try build successful - checks-actions, checks-azure |
|
Queued af826c3de514b426f1172de2883db4a251e45c65 with parent 8fe73e8, future comparison URL. |
|
Finished benchmarking try commit (af826c3de514b426f1172de2883db4a251e45c65): comparison url. Benchmarking this pull request likely means that it is perf-sensitive, so we're automatically marking it as not fit for rolling up. Please note that if the perf results are neutral, you should likely undo the rollup=never given below by specifying Importantly, though, if the results of this run are non-neutral do not roll this PR up -- it will mask other regressions or improvements in the roll up. @bors rollup=never |
| match spec.target_spec().arch.as_str() { | ||
| // System-V will pass return values up to 128 bits in RAX/RDX. | ||
| "x86_64" => Size::from_bits(128), |
There was a problem hiding this comment.
I was hoping removing this string comparison would help perf but I guess not?
| if is_ret { call::max_ret_by_val(cx) } else { Pointer.size(cx) }; | ||
| // Return structures up to 2 pointers in size by value, matching `ScalarPair`. LLVM | ||
| // will usually return these in 2 registers, which is more efficient than by-ref. | ||
| let max_by_val_size = if is_ret { Pointer.size(cx) * 2 } else { Pointer.size(cx) }; |
There was a problem hiding this comment.
I wonder if it makes sense to limit this to returns, but also whether 3 might work better than 2 as the multiplier.
But I guess there's a risk it might get downgraded from registers... I really wish LLVM modeled ABIs more precisely so we can control that.
|
@bors r+ |
|
📌 Commit b01694e has been approved by |
bors
added
S-waiting-on-bors
|
☀️ Test successful - checks-actions, checks-azure |
Pass arguments up to 2*usize by value In rust-lang#77434 (comment), `@eddyb` said: > I wonder if it makes sense to limit this to returns [...] Let's do a perf run and find out. It seems the `extern "C"` ABI will pass arguments up to 2*usize in registers: https://godbolt.org/z/n8E6zc. (modified from rust-lang#26494 (comment)) r? `@nagisa`
Addresses #76986 (comment) and #76986 (comment) by doing the optimization on all targets.
This matches what we do for functions returning
&[T]and other fat pointers, so it should be Harmless™