Change ArrayBase.ptr to NonNull type

[berquist]

Closes#434

Should any of these unwrap calls turn into proper error handling? It looks like most functions already state that they'll panic if given an invalid pointer.

It also looks like there's some wasteful *mut A -> NonNull<A> -> *mut A or NonNull<A> -> *mut A -> NonNull<A> conversions happening (for example, in impl_clone). What do you think about eliminating these?

[bluss]

A fun very loose concern (I'm sorry, might just be interesting!)

Rust has in the past had problems with loop optimization of iterators that yield Option<&T> because of the non-nullable pointer optimization that is used for references in Option. Long story short, the reference compiles to just a pointer, there are some attributes that tell the compiler the pointer will be non-null, and simultaneously the code will use the null value to encode Option::None.

Then the non-nullable annotation can get lost or confused in some optimization passes, and the result is that it's not always certain if null means the end of the loop or just the next pointer result of the iteration.

When using NonNull inside ArrayView like this, it's possible that the same misoptimization can show up in Iterators yielding array views because Option<ArrayView<..>> could have the same problem.

Wait - but don't the most important iterators in Rust use Option<&T>, like the slice, so it must be good? They are in fact tuned with assume etc: https://doc.rust-lang.org/1.37.0/src/core/slice/mod.rs.html#525https://doc.rust-lang.org/1.37.0/src/core/slice/mod.rs.html#3102 which we don't have access to in stable Rust.

We can just look out for it and check benchmarks

[berquist]

Alright, I think I got it.

Some benchmarks:

• master branch
• this branch, first version
• this branch, after eliminating round-tripping in data_traits
• this branch, after doing unchecked NonNull creation

It looks like there might be a regression in taking dot products.

[bluss]

Thanks a lot, I'd love to take a closer look at this. This is a good tool for benchmark comparisons: https://github.com/BurntSushi/cargo-benchcmp

From master to the latest after unchecked:

 name 63 ns/iter 62 ns/iter diff ns/iter diff % add_1d_regular 369 373 4 1.08% add_1d_strided 1,449 1,481 32 2.21% add_2d_0_to_2_iadd_scalar 247 252 5 2.02% add_2d_alloc 699 711 12 1.72% add_2d_assign_ops 374 382 8 2.14% add_2d_both_transposed 1,343 1,321 -22 -1.64% add_2d_broadcast_0_to_2 248 251 3 1.21% add_2d_broadcast_1_to_2 902 480 -422 -46.78% add_2d_cutout 924 547 -377 -40.80% add_2d_cutouts_by_16 1,609 1,158 -451 -28.03% add_2d_cutouts_by_32 1,169 753 -416 -35.59% add_2d_cutouts_by_4 6,105 5,303 -802 -13.14% add_2d_f32_regular 376 380 4 1.06% add_2d_one_transposed 1,200 1,266 66 5.50% add_2d_regular 376 378 2 0.53% add_2d_regular_dyn 428 431 3 0.70% add_2d_strided 1,274 1,370 96 7.54% add_2d_strided_dyn 1,278 1,378 100 7.82% add_2d_zip 872 392 -480 -55.05% add_2d_zip_alloc 491 496 5 1.02% add_2d_zip_both_transposed 875 394 -481 -54.97% add_2d_zip_cutout 967 628 -339 -35.06% add_2d_zip_one_transposed 1,188 1,271 83 6.99% add_2d_zip_strided 1,272 1,362 90 7.08% add_3d_strided 2,520 2,361 -159 -6.31% add_3d_strided_dyn 2,525 2,346 -179 -7.09% assign_scalar_2d_corder 483 492 9 1.86% assign_scalar_2d_cutout 995 1,007 12 1.21% assign_scalar_2d_forder 488 494 6 1.23% assign_zero_2d_corder 483 499 16 3.31% assign_zero_2d_cutout 995 1,017 22 2.21% assign_zero_2d_forder 488 499 11 2.25% bench_col_iter 367 369 2 0.54% bench_iter_diag 486 255 -231 -47.53% bench_row_iter 252 252 0 0.00% bench_to_owned_n 90 93 3 3.33% bench_to_owned_strided 349 363 14 4.01% bench_to_owned_t 90 93 3 3.33% chunk2x2_iter_sum 307,793 308,089 296 0.10% chunk2x2_sum 132,500 129,286 -3,214 -2.43% chunk2x2_sum_get1 139,509 138,343 -1,166 -0.84% chunk2x2_sum_get2 143,527 142,770 -757 -0.53% chunk2x2_sum_uget1 104,488 104,178 -310 -0.30% clip 400 388 -12 -3.00% create_iter_4d 22 22 0 0.00% default_f64 1,999 2,014 15 0.75% dot_extended 1,892 1,776 -116 -6.13% dot_f32_1024 88 92 4 4.55% dot_f32_10e6 90,033 91,441 1,408 1.56% dot_f32_16 4 4 0 0.00% dot_f32_20 5 5 0 0.00% dot_f32_256 22 22 0 0.00% dot_f32_32 4 5 1 25.00% equality_f32 499 498 -1 -0.20% equality_f32_mixorder 2,737 2,870 133 4.86% equality_i32 496 505 9 1.81% fold_axis 554 538 -16 -2.89% fold_sum_i32_2d_cutout 343 300 -43 -12.54% fold_sum_i32_2d_cutout_transpose 343 301 -42 -12.24% fold_sum_i32_2d_regular 138 139 1 0.72% fold_sum_i32_2d_stride 978 997 19 1.94% fold_sum_i32_2d_transpose 139 143 4 2.88% gemv_64_32 585 590 5 0.85% gemv_64_64c 864 852 -12 -1.39% gemv_64_64f 2,176 2,193 17 0.78% iadd_scalar_2d_regular 248 252 4 1.61% iadd_scalar_2d_regular_dyn 281 279 -2 -0.71% iadd_scalar_2d_strided 1,066 1,075 9 0.84% iadd_scalar_2d_strided_dyn 1,050 1,077 27 2.57% indexed_iter_1d_ix1 1,424 965 -459 -32.23% indexed_iter_2d_ix2 3,797 2,605 -1,192 -31.39% indexed_iter_3d_dyn 111,694 111,813 119 0.11% indexed_iter_3d_ix3 5,892 3,025 -2,867 -48.66% indexed_zip_1d_ix1 973 969 -4 -0.41% indexed_zip_2d_ix2 2,653 1,970 -683 -25.74% indexed_zip_3d_ix3 2,922 2,913 -9 -0.31% iter_all_2d_cutout 8,461 8,499 38 0.45% iter_axis_chunks_1_iter_sum 1,452 1,468 16 1.10% iter_axis_chunks_5_iter_sum 604 614 10 1.66% iter_axis_iter_sum 534 509 -25 -4.68% iter_filter_sum_2d_f32 310 314 4 1.29% iter_filter_sum_2d_stride_f32 167 167 0 0.00% iter_filter_sum_2d_stride_u32 108 108 0 0.00% iter_filter_sum_2d_u32 38 37 -1 -2.63% iter_sum_1d_raw 137 139 2 1.46% iter_sum_1d_regular 137 139 2 1.46% iter_sum_1d_strided_fold 1,219 1,205 -14 -1.15% iter_sum_1d_strided_rfold 1,212 1,199 -13 -1.07% iter_sum_2d_by_row 277 259 -18 -6.50% iter_sum_2d_cutout 396 400 4 1.01% iter_sum_2d_cutout_by_row 358 355 -3 -0.84% iter_sum_2d_cutout_outer_iter 278 260 -18 -6.47% iter_sum_2d_raw 138 138 0 0.00% iter_sum_2d_regular 143 143 0 0.00% iter_sum_2d_transpose 1,119 1,117 -2 -0.18% iter_sum_2d_transpose_by_row 1,979 1,858 -121 -6.11% iter_sum_2d_transpose_regular 3,741 4,392 651 17.40% map_axis_0 465 465 0 0.00% map_axis_1 416 489 73 17.55% map_regular 30 30 0 0.00% map_stride 85 87 2 2.35% map_stride_double_f64 334 304 -30 -8.98% map_stride_f64 506 510 4 0.79% map_stride_u32 524 535 11 2.10% mat_mul_f32::m004 203 223 20 9.85% mat_mul_f32::m007 258 257 -1 -0.39% mat_mul_f32::m008 214 198 -16 -7.48% mat_mul_f32::m012 575 577 2 0.35% mat_mul_f32::m016 584 569 -15 -2.57% mat_mul_f32::m032 3,105 3,137 32 1.03% mat_mul_f32::m064 21,492 21,665 173 0.80% mat_mul_f32::m127 162,131 163,678 1,547 0.95% mat_mul_f32::mix10000 12,692,607 12,809,887 117,280 0.92% mat_mul_f32::mix16x4 691 709 18 2.60% mat_mul_f32::mix32x2 508 506 -2 -0.39% mat_mul_f64::m004 195 218 23 11.79% mat_mul_f64::m007 282 342 60 21.28% mat_mul_f64::m008 237 274 37 15.61% mat_mul_f64::m012 591 646 55 9.31% mat_mul_f64::m016 773 827 54 6.99% mat_mul_f64::m032 4,432 4,435 3 0.07% mat_mul_f64::m064 31,081 31,231 150 0.48% mat_mul_f64::m127 235,004 235,987 983 0.42% mat_mul_f64::mix10000 18,528,812 18,515,705 -13,107 -0.07% mat_mul_f64::mix16x4 917 976 59 6.43% mat_mul_f64::mix32x2 719 762 43 5.98% mat_mul_i32::m004 130 128 -2 -1.54% mat_mul_i32::m007 402 403 1 0.25% mat_mul_i32::m008 553 562 9 1.63% mat_mul_i32::m012 1,299 1,334 35 2.69% mat_mul_i32::m016 2,600 2,579 -21 -0.81% mat_mul_i32::m032 14,575 14,584 9 0.06% mat_mul_i32::m064 92,357 97,458 5,101 5.52% mat_mul_i32::m127 780,099 785,692 5,593 0.72% mean_axis0 2,755 2,717 -38 -1.38% mean_axis1 2,283 2,212 -71 -3.11% scalar_add_1 547 577 30 5.48% scalar_add_2 550 577 27 4.91% scalar_sub_1 549 576 27 4.92% scalar_sub_2 573 603 30 5.24% scalar_toowned 328 360 32 9.76% scaled_add_2d_f32_regular 379 379 0 0.00% sum_2d_cutout 372 373 1 0.27% sum_2d_float 362 362 0 0.00% sum_2d_float_cutout 419 420 1 0.24% sum_2d_float_t_cutout 1,499 1,496 -3 -0.20% sum_2d_regular 212 210 -2 -0.94% sum_3_azip 4,738 4,733 -5 -0.11% sum_3_azip_fold 1,012 1,006 -6 -0.59% sum_3_std_zip1 1,007 1,000 -7 -0.70% sum_3_std_zip2 1,008 991 -17 -1.69% sum_3_std_zip3 1,013 991 -22 -2.17% sum_axis0 2,746 2,713 -33 -1.20% sum_axis1 2,318 2,196 -122 -5.26% vector_sum_3_azip 2,438 2,396 -42 -1.72% vector_sum_3_std_zip 2,406 2,408 2 0.08% vector_sum_3_zip_unchecked 2,418 2,394 -24 -0.99% vector_sum_3_zip_unchecked_manual 2,440 2,405 -35 -1.43% zeros_f64 2,020 2,029 9 0.45% 

[bluss]

I've used the breaking-change label so that we remember to put a note in the relnotes, that raw pointers are now checked to be non-null and can never be null, even in empty array views (This has been true before too, but it's true now too 🙂).

[bluss]

I've pushed two commits to the PR

One adds dep on rawpointer 0.2 which has added an .offset() method for NonNull, which makes our code a lot simpler. (See commit log for more info).
I thought adding this method there made sense, and it was already a transitive dep through matrixmultiply — although that's the 0.1 version, but matrixmultiply will update too.

We could also copy the offset method into this crate, to reduce deps.

Using .offset() simplifies a lot, and I factored out the two other way nonnull pointers were created.

I hope this is ok. You've enabled maintainer pushes to the PR, and that's the reason I could update your branch.

[bluss]

I'm sorry for the mess. I've merged master and resolved conflicts, since another PR made this one conflict. (We could also rebase - please rebase and squash as needed if you want — I thought it was awkward to do that in a two-author branch).

[bluss]

I think this is ready for merge, but it could be rebase/squashed if you want @berquist

[berquist]

Whatever you would prefer.

[jturner314]

Great work on this! There are some places where it would be nice to replace more raw pointers with NonNull (e.g. in the iterators), but this PR is a definite improvement.