str::is_char_boundary - slight optimization

[Soveu]

Current str::is_char_boundary implementation emits slightly more instructions, because it includes an additional branch for index == s.len()

pubfnis_char_boundary(s:&str,index:usize) -> bool{if index == 0 || index == s.len(){returntrue;}match s.as_bytes().get(index){None => false,Some(&b) => (b asi8) >= -0x40,}}

Just changing the place of index == s.len() merges it with index < s.len() from s.as_bytes().get(index)

pubfnis_char_boundary2(s:&str,index:usize) -> bool{if index == 0{returntrue;}match s.as_bytes().get(index){// For some reason, LLVM likes this comparison here moreNone => index == s.len(),// This is bit magic equivalent to: b < 128 || b >= 192Some(&b) => (b asi8) >= -0x40,}}

This one has better codegen on every platform, except powerpc

x86 codegen

example::is_char_boundary:moval,1testrdx,rdxje .LBB0_5cmprsi,rdxje .LBB0_5cmprsi,rdxjbe .LBB0_3cmp byte ptr [rdi+rdx],-65 setg al.LBB0_5:ret.LBB0_3:xoreax,eaxretexample::is_char_boundary2:testrdx,rdxje .LBB1_1cmprsi,rdxjbe .LBB1_4cmp byte ptr [rdi+rdx],-65 setg alret.LBB1_1: ; technically this branch is the same as LBB1_4moval,1ret.LBB1_4: sete alret

aarch64 codegen

example::is_char_boundary:
mov x8, x0
mov w0, #1
cbz x2, .LBB0_4
cmp x1, x2
b.eq .LBB0_4b.ls .LBB0_5
ldrsb w8, [x8, x2]
cmn w8, #65
cset w0, gt
.LBB0_4:
ret
.LBB0_5:
mov w0, wzr
ret
example::is_char_boundary2:
cbz x2, .LBB1_3
cmp x1, x2
b.ls .LBB1_4
ldrsb w8, [x0, x2]
cmn w8, #65
cset w0, gt
ret
.LBB1_3:
mov w0, #1
ret
.LBB1_4:
cset w0, eq
ret

riscv64gc codegen

example::is_char_boundary:
seqz a3, a2
xor a4, a1, a2
seqz a4, a4
or a4, a4, a3
addi a3, zero, 1
bnez a4, .LBB0_3
bgeu a2, a1, .LBB0_4
add a0, a0, a2
lb a0, 0(a0)
addi a1, zero, -65
slt a3, a1, a0
.LBB0_3:
mv a0, a3
ret
.LBB0_4:
mv a0, zero
ret

example::is_char_boundary2:
beqz a2, .LBB1_3
bgeu a2, a1, .LBB1_4
add a0, a0, a2
lb a0, 0(a0)
addi a1, zero, -65
slt a0, a1, a0
ret
.LBB1_3:
addi a0, zero, 1
ret
.LBB1_4:
xor a0, a1, a2
seqz a0, a0
ret

Link to godbolt

@rustbot label: A-codegen

[rust-highfive]

r? @yaahc

(rust-highfive has picked a reviewer for you, use r? to override)

[jyn514]

@Soveu fyi - A-codegen is usually used for the codegen part of the compiler itself, not for changes to the standard library that affect performance.

[Soveu]

btw, can I add I-slow label myself or only members can do it?

[jyn514]

You should be able to - try pinging rustbot with label: +I-slow?

[Soveu]

I was thinking all the time that only members can add those fancy red I-something labels 👀
@rustbot label: +I-slow

[klensy]

Slightly less cryptic version https://godbolt.org/z/Y9o86ch6h

pubfnis_char_boundary(s:&str,index:usize) -> bool{match s.as_bytes().get(index){None => false,Some(_)if index == 0 || index == s.len() => true,Some(&b) => (b asi8) >= -0x40,}}

but this swaps order of cmp and test, better benchmark all versions to check actual difference.

[Soveu]

Slightly less cryptic version https://godbolt.org/z/Y9o86ch6h

pubfnis_char_boundary(s:&str,index:usize) -> bool{match s.as_bytes().get(index){None => false,Some(_)if index == 0 || index == s.len() => true,Some(&b) => (b asi8) >= -0x40,}}

but this swaps order of cmp and test, better benchmark all versions to check actual difference.

index == s.len() will never fire if get(index) returned Some

[klensy]

index == s.len() will never fire if get(index) returned Some

Upps

pubfnis_char_boundary(s:&str,index:usize) -> bool{match s.as_bytes().get(index){Noneif index == s.len() => true,None => false,Some(_)if index == 0 => true,Some(&b) => (b asi8) >= -0x40,}}

But this looks bad.

[Soveu]

I will soon add some comments

[Soveu]

Plus, I want to test how it affects s.get(..i) and s.get(i..) which rely heavily on returning early if index == 0 or index == s.len() (and LLVM actually optimizing it)

[Soveu]

Both on opt-level=3 and opt-level=2 there is only one bound check in get(..i) and get(i..).
On opt-level=1 it's more complicated:

• functions aren't inlined by default, so no cross-function optimizations can occur
• when forcing it via #[inline(always)], the new implementation loses at get(i..), because the index == s.len() is "hidden" after s.as_bytes().get(i)

[klensy]

pubfnis_char_boundary(s:&str,index:usize) -> bool{if index == 0{true}elseif index < s.len(){unsafe{*s.as_bytes().get_unchecked(index)asi8 >= -0x40}}elseif index == s.len(){true}else{false}}

[Soveu]

@yaahc ?

[Amanieu]

@bors try @rust-timer queue

[rust-timer]

Awaiting bors try build completion.

@rustbot label: +S-waiting-on-perf

[bors]

⌛ Trying commit 7bd9d9f with merge b8c3fd3abbd324a8ce3163e2d72ad2da5bb38d83...

[bors]

☀️ Try build successful - checks-actions
Build commit: b8c3fd3abbd324a8ce3163e2d72ad2da5bb38d83 (b8c3fd3abbd324a8ce3163e2d72ad2da5bb38d83)

[rust-timer]

Queued b8c3fd3abbd324a8ce3163e2d72ad2da5bb38d83 with parent 69b352e, future comparison URL.

[rust-timer]

Finished benchmarking try commit (b8c3fd3abbd324a8ce3163e2d72ad2da5bb38d83): 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 rollup- to bors.

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
@rustbot label: +S-waiting-on-review -S-waiting-on-perf

[Amanieu]

Perf changes seem to be in the noise.

@bors r+ rollup=maybe

[bors]

📌 Commit 7bd9d9f has been approved by Amanieu

[Soveu]

Perf changes seem to be in the noise.

I was kinda expecting it, it's not a huge thing. What about binary size?

[matthieu-m]

@klensy I'm surprised by those constructs:

Noneif index == s.len() => true,None => false,

if index == s.len(){true}else{false}

Naively, I would expect to just return index == s.len() rather branching on it to select true or false.

Is there any particular reason for the unusual branch?

[klensy]

@matthieu-m

All that if branches is straightforward walk over index, starting from 0 and up, checking current state, but don't require to think about None => index == self.len() trick. Plus, that compiles to exactly the same asm, if i remember that correctly.