Eliminate excessive null-checks from slice iterators

[dotdash]

The data pointer used in the slice is never null, using assume() to tell
LLVM about it gets rid of various unneeded null checks when iterating
over the slice.

Since the snapshot compiler is still using an older LLVM version, omit
the call in stage0, because compile times explode otherwise.

Benchmarks from #18193

running 5 tests
test _range ... bench: 33329 ns/iter (+/- 417)
test assembly ... bench: 33299 ns/iter (+/- 58)
test enumerate ... bench: 33318 ns/iter (+/- 83)
test iter ... bench: 33311 ns/iter (+/- 130)
test position ... bench: 33300 ns/iter (+/- 47)
test result: ok. 0 passed; 0 failed; 0 ignored; 5 measured

Fixes#18193

[rust-highfive]

r? @nikomatsakis

(rust_highfive has picked a reviewer for you, use r? to override)

[huonw]

From the description I take it this doesn't suffer from the compile time problem mentioned in #21418 (comment)? (Out of interest what changed in LLVM? Just general optimisations?)

Also, I found that the compiler aborted compiling libcore (can't remember if it was stage1 or stage2) when I did a similar change (#21448); this doesn't suffer from that?

[dotdash]

From the description I take it this doesn't suffer from the compile time problem mentioned in #21418 (comment)?

Right, I compared compile times for (IIRC) rustc_trans and rustc_typeck and there was no measureable difference in compile times with the assume in stage 2 (or maybe it was stage 1 that I tested, shouldn't matter).

I also just bootstrapped a new version of #21418 and that took ~24 minutes from finishing stage0 libcore to finishing stage2 librustc. Considering that just rustc_trans took 13 minutes when the slowdown was there, it looks good to me ;-)

(Out of interest what changed in LLVM? Just general optimisations?)

I don't know.

Also, I found that the compiler aborted compiling libcore (can't remember if it was stage1 or stage2) when I did a similar change (#21448); this doesn't suffer from that?

The benchmark was taken with the stage2 rustc, so no, it didn't abort.

[huonw]

Could you add the assume to next_back as well?

[dotdash]

Why next_back? Wouldn't as_mut_slice and into_iter be the places that correspond to as_slice?

I originally had it in next (which is why I forgot to add it to as_mut_slice, because next is generated in a macro that is used for both iter types), and that gave worse results, because the assumption got lost at some point when SROA (I think) replaced the load in next() with the one from as_slice.

[huonw]

Oh, huh. I totally misread this PR. I was thinking this was in <slice::Iter<T> as Iterator>::next.

[huonw]

I suspect that means that this doesn't solve a case like http://is.gd/pteFUM,

.LBB0_2: testq %rdx, %rdxje .LBB0_4 addl (%rdx), %eax addq$4, %rdx addq $-4, %rcxjne .LBB0_2

since there's no Vec involved. I wonder if that would be addressed by placing assumptions in the iter/iter_mut constructor?

[dotdash]

Will try to see if that helps

[jdm]

ptr::null_mut() here and all other uses of 0 as *mut T

[dotdash]

Actually, using is_null() works now, that didn't work the first time I tried, either because of an error on my side or because of the ptrtoint instruction it was generated with the old is_null() implementation.
Thanks!

[dotdash]

I wonder if that would be addressed by placing assumptions in the iter/iter_mut constructor?

At least in this case, the assume() gets canonicalized to !nonnull metadata on a load, and that gets lost in the SROA pass. I've added the call anyway, hoping for it to be useful in the future.

[huonw]

@bors r+ e49f

[alexcrichton]

@bors: rollup

[alexcrichton]

@bors: rollup-

[bors]

⌛ Testing commit e49f6d6 with merge cd0d956...

[bors]

💔 Test failed - auto-win-32-nopt-t

[bluss]

@huonw, that fold testcase seems to have been resolved now, it vectorizes! (rustc 1.0.0-nightly (ba2f13ef0 2015-02-04 20:03:55 +0000) Probably due to the other awesome nonnull changes?

[huonw]

The failure seems legitimate; although, possibly a LLVM bug?

[dotdash]

Yeah, I can reproduce that in my Windows VM.

[dotdash]

Yup, LLVM bug: http://reviews.llvm.org/D7533

[dotdash]

@bors r=huonw 7412d1b

[bors]

⌛ Testing commit 7412d1b with merge c103604...

[bors]

💔 Test failed - auto-win-32-nopt-t

[alexcrichton]

@bors: retry

[bors]

⌛ Testing commit 7412d1b with merge a1cfc62...

[bors]

💔 Test failed - auto-win-32-nopt-t

[aturon]

@bors: retry

[frol]

Could someone, please, point me to a tracking issue or a documentation about the excessive null-checks?

I want to learn more about the cause of extra testq %rsi, %rsi (test rsi, rsi) blocks when I compile the following snippets:

pubfnfoo(s:&[u32]) -> u32{
s.iter().sum()}

pubfnfoo(s:&[u32]) -> u32{
s.iter().sum()}

pubfnfoo(s:Vec<i32>) -> usize{
s.len()}

and even

pubfnfoo(_s:Vec<i32>) -> usize{0}

produces assembly code which might have been avoided:

playground::foo:movq8(%rdi), %rsi	testq	%rsi, %rsije	.LBB0_2	pushq	%raxmovq	(%rdi), %rdi	shlq$2, %rsi	movl$4, %edx	callq	__rust_dealloc@PLT	addq$8, %rsp.LBB0_2:	xorl	%eax, %eax	retq

[nagisa]

In both these cases the length, rather than pointer is being checked. For the slice case as early exit before a vectorised solution is used, for Vec drop case because Vec may actually not have backing storage (Vec::new does not allocate).

[frol]

for Vec drop case because Vec may actually not have backing storage (Vec::new does not allocate).

Indeed, I completely forgot that the owned object must be deallocated here.

For the slice case as early exit before a vectorised solution is used

Right...

@nagisa I am sorry for the trouble! Rust is awesome! :)