Make the arena fast path smaller

[Zoxc]

r? @michaelwoerister

[Zoxc]

@bors try

[bors]

⌛ Trying commit 45da76e30802c4e272c6e82293d41142a2a8047f with merge b40fd9ad1c114049c0ef02dd1a36e93788ec0482...

[bors]

☀️ Test successful - status-travis
State: approved= try=True

[Zoxc]

@rust-timer build b40fd9ad1c114049c0ef02dd1a36e93788ec0482

[rust-timer]

Success: Queued b40fd9ad1c114049c0ef02dd1a36e93788ec0482 with parent d3ed348, comparison URL.

[rust-timer]

Finished benchmarking try commit b40fd9ad1c114049c0ef02dd1a36e93788ec0482

[michaelwoerister]

The perf-results mostly look like noise (except for ctfe-stress). I'm not sure this is worth the more complicated implementation.

cc @rust-lang/wg-compiler-performance

[Zoxc]

The results doesn't look like noise though. The whole point of the arena is performance. We should just use malloc otherwise =P

[Zoxc]

This is also a nice size optimization. It brings the fast path for allocating an usize from:

.text:0000000180001235                 mov     rsi, rcx
.text:0000000180001238                 mov     rax, [rcx]
.text:000000018000123B                 mov     rcx, [rcx+8]
.text:000000018000123F                 add     rax, 7
.text:0000000180001243                 and     rax, 0FFFFFFFFFFFFFFF8h
.text:0000000180001247                 mov     [rsi], rax
.text:000000018000124A                 cmp     rcx, rax
.text:000000018000124D                 jb      short loc_180001281
.text:000000018000124F                 mov     rdx, rax
.text:0000000180001252                 add     rdx, 8
.text:0000000180001256                 cmp     rdx, rcx
.text:0000000180001259                 jnb     short loc_18000126B
.text:000000018000125B
.text:000000018000125B loc_18000125B:                          ; CODE XREF: alloctest+4F↓j
.text:000000018000125B                 mov     [rsi], rdx
...
.text:000000018000126B ; ---------------------------------------------------------------------------
.text:000000018000126B
.text:000000018000126B loc_18000126B:                          ; CODE XREF: alloctest+29↑j
.text:000000018000126B                 mov     edx, 8
.text:0000000180001270                 mov     rcx, rsi
.text:0000000180001273                 call    _ZN5arena13DroplessArena4grow17h2cc6dfe0f3889a5eE ; arena::DroplessArena::grow::h2cc6dfe0f3889a5e
.text:0000000180001278                 mov     rax, [rsi]
.text:000000018000127B                 lea     rdx, [rax+8]
.text:000000018000127F                 jmp     short loc_18000125B
.text:0000000180001281 ; ---------------------------------------------------------------------------
.text:0000000180001281
.text:0000000180001281 loc_180001281:                          ; CODE XREF: alloctest+1D↑j
.text:0000000180001281                 lea     rcx, aAssertionFaile ; "assertion failed: self.ptr <= self.end"
.text:0000000180001288                 lea     r8, off_180003230
.text:000000018000128F                 mov     edx, 26h
.text:0000000180001294                 call    _ZN3std9panicking11begin_panic17hca0ac606c2739ec6E ; std::panicking::begin_panic::hca0ac606c2739ec6
.text:0000000180001299 ; ---------------------------------------------------------------------------
.text:0000000180001299                 ud2
.text:0000000180001299 alloctest       endp

to:

.text:00000001800012C4                 mov     rax, [rcx]
.text:00000001800012C7                 lea     rdx, [rax+8]
.text:00000001800012CB                 cmp     rdx, [rcx+8]
.text:00000001800012CF                 jnb     short loc_1800012E0
.text:00000001800012D1                 mov     [rcx], rdx
.text:00000001800012D4
.text:00000001800012D4 loc_1800012D4:                          ; CODE XREF: alloctest+2A↓j
...
.text:00000001800012E0 ; ---------------------------------------------------------------------------
.text:00000001800012E0
.text:00000001800012E0 loc_1800012E0:                          ; CODE XREF: alloctest+F↑j
.text:00000001800012E0                 mov     edx, 8
.text:00000001800012E5                 call    _ZN5arena13DroplessArena18grow_and_alloc_raw17ha43c166c6d43e92fE ; arena::DroplessArena::grow_and_alloc_raw::ha43c166c6d43e92f
.text:00000001800012EA                 jmp     short loc_1800012D4

[michaelwoerister]

I can see the first change making sense but why change from u8 to usize for the drop-less arena?

[Zoxc]

@michaelwoerister It avoids the instructions realigning the bump pointer by keeping it always usize aligned.

[Zoxc]

Might make the DroplessArena generic over the backing type, so we could use u8 in the string interner. I think all other allocations are usize aligned.

[michaelwoerister]

I'm still not convinced that there are performance gains here. Compare this to a perf.rlo run from a few days ago where I accidentally benchmarked to equivalent versions of the compiler (so we know what we see is noise):
https://perf.rust-lang.org/compare.html?start=10e2c729ea2e85cb1e2a08be40564492f49e45ec&end=5280d23d48ea46ca2469a77cc240fbef6767efbb

Except for ctfe-stress it looks the same to me.

[nnethercote]

I agree that the result are probably just noise, with the possible exception of ctfe-stress, but it is one of the most variable benchmarks, which is why it has the '?' markers on its results. So even that could be noise.

[michaelwoerister]

If you remove the BackingType change, I'm happy to approve the PR. I think that always using u8 as the backing type makes the code a lot easier to reason about.