Mercurial > public > mercurial-scm > hg-stable
view rust/hg-core/src/revlog/patch.rs @ 46155:fce2f20a54ce
copies-rust: start recording overwrite as they happens
If a revision has information overwriting data from another revision, the
overwriting revision is a descendant of the overwritten one. So we could warm
the Oracle cache with such information to avoid potential future `is_ancestors`
call.
This provide us with a large speedup in the most expensive cases:
Repo Case Source-Rev Dest-Rev # of revisions old time new time Difference Factor time per rev
---------------------------------------------------------------------------------------------------------------------------------------------------------------
mozilla-try x00000_revs_x00000_added_x0000_copies 1b661134e2ca 1ae03d022d6d : 228985 revs, 41.113063 s, 36.001255 s, -5.111808 s, ? 0.8757, 157 ?s/rev
mozilla-try x00000_revs_x00000_added_x000_copies 9b2a99adc05e 8e29777b48e6 : 382065 revs, 27.891612 s, 14.340641 s, -13.550971 s, ? 0.5142, 37 ?s/rev
Full comparison below:
Repo Case Source-Rev Dest-Rev # of revisions old time new time Difference Factor time per rev
---------------------------------------------------------------------------------------------------------------------------------------------------------------
mercurial x_revs_x_added_0_copies ad6b123de1c7 39cfcef4f463 : 1 revs, 0.000042 s, 0.000042 s, +0.000000 s, ? 1.0000, 42 ?s/rev
mercurial x_revs_x_added_x_copies 2b1c78674230 0c1d10351869 : 6 revs, 0.000114 s, 0.000109 s, -0.000005 s, ? 0.9561, 18 ?s/rev
mercurial x000_revs_x000_added_x_copies 81f8ff2a9bf2 dd3267698d84 : 1032 revs, 0.004934 s, 0.004953 s, +0.000019 s, ? 1.0039, 4 ?s/rev
pypy x_revs_x_added_0_copies aed021ee8ae8 099ed31b181b : 9 revs, 0.000195 s, 0.000237 s, +0.000042 s, ? 1.2154, 26 ?s/rev
pypy x_revs_x000_added_0_copies 4aa4e1f8e19a 359343b9ac0e : 1 revs, 0.000050 s, 0.000050 s, +0.000000 s, ? 1.0000, 50 ?s/rev
pypy x_revs_x_added_x_copies ac52eb7bbbb0 72e022663155 : 7 revs, 0.000113 s, 0.000113 s, +0.000000 s, ? 1.0000, 16 ?s/rev
pypy x_revs_x00_added_x_copies c3b14617fbd7 ace7255d9a26 : 1 revs, 0.6f1f4a s, 0.6f1f4a s, +0.000000 s, ? 1.0000, 322 ?s/rev
pypy x_revs_x000_added_x000_copies df6f7a526b60 a83dc6a2d56f : 6 revs, 0.010788 s, 0.010702 s, -0.000086 s, ? 0.9920, 1783 ?s/rev
pypy x000_revs_xx00_added_0_copies 89a76aede314 2f22446ff07e : 4785 revs, 0.050880 s, 0.050504 s, -0.000376 s, ? 0.9926, 10 ?s/rev
pypy x000_revs_x000_added_x_copies 8a3b5bfd266e 2c68e87c3efe : 6780 revs, 0.081760 s, 0.080159 s, -0.001601 s, ? 0.9804, 11 ?s/rev
pypy x000_revs_x000_added_x000_copies 89a76aede314 7b3dda341c84 : 5441 revs, 0.061382 s, 0.060058 s, -0.001324 s, ? 0.9784, 11 ?s/rev
pypy x0000_revs_x_added_0_copies d1defd0dc478 c9cb1334cc78 : 43645 revs, 0.585802 s, 0.536950 s, -0.048852 s, ? 0.9166, 12 ?s/rev
pypy x0000_revs_xx000_added_0_copies bf2c629d0071 4ffed77c095c : 2 revs, 0.012803 s, 0.012868 s, +0.000065 s, ? 1.0051, 6434 ?s/rev
pypy x0000_revs_xx000_added_x000_copies 08ea3258278e d9fa043f30c0 : 11316 revs, 0.113558 s, 0.112806 s, -0.000752 s, ? 0.9934, 9 ?s/rev
netbeans x_revs_x_added_0_copies fb0955ffcbcd a01e9239f9e7 : 2 revs, 0.000085 s, 0.000084 s, -0.000001 s, ? 0.9882, 42 ?s/rev
netbeans x_revs_x000_added_0_copies 6f360122949f 20eb231cc7d0 : 2 revs, 0.000106 s, 0.000106 s, +0.000000 s, ? 1.0000, 53 ?s/rev
netbeans x_revs_x_added_x_copies 1ada3faf6fb6 5a39d12eecf4 : 3 revs, 0.000175 s, 0.000174 s, -0.000001 s, ? 0.9943, 58 ?s/rev
netbeans x_revs_x00_added_x_copies 35be93ba1e2c 9eec5e90c05f : 9 revs, 0.000721 s, 0.000726 s, +0.000005 s, ? 1.0069, 80 ?s/rev
netbeans x000_revs_xx00_added_0_copies eac3045b4fdd 51d4ae7f1290 : 1421 revs, 0.010127 s, 0.010105 s, -0.000022 s, ? 0.9978, 7 ?s/rev
netbeans x000_revs_x000_added_x_copies e2063d266acd 6081d72689dc : 1533 revs, 0.015616 s, 0.015748 s, +0.000132 s, ? 1.0085, 10 ?s/rev
netbeans x000_revs_x000_added_x000_copies ff453e9fee32 411350406ec2 : 5750 revs, 0.061341 s, 0.060357 s, -0.000984 s, ? 0.9840, 10 ?s/rev
netbeans x0000_revs_xx000_added_x000_copies 588c2d1ced70 1aad62e59ddd : 66949 revs, 0.542214 s, 0.499356 s, -0.042858 s, ? 0.9210, 7 ?s/rev
mozilla-central x_revs_x_added_0_copies 3697f962bb7b 7015fcdd43a2 : 2 revs, 0.000089 s, 0.000092 s, +0.000003 s, ? 1.0337, 46 ?s/rev
mozilla-central x_revs_x000_added_0_copies dd390860c6c9 40d0c5bed75d : 8 revs, 0.000279 s, 0.000279 s, +0.000000 s, ? 1.0000, 34 ?s/rev
mozilla-central x_revs_x_added_x_copies 8d198483ae3b 14207ffc2b2f : 9 revs, 0.000184 s, 0.000186 s, +0.000002 s, ? 1.0109, 20 ?s/rev
mozilla-central x_revs_x00_added_x_copies 98cbc58cc6bc 446a150332c3 : 7 revs, 0.000661 s, 0.000660 s, -0.000001 s, ? 0.9985, 94 ?s/rev
mozilla-central x_revs_x000_added_x000_copies 3c684b4b8f68 0a5e72d1b479 : 3 revs, 0.003377 s, 0.003372 s, -0.000005 s, ? 0.9985, 1124 ?s/rev
mozilla-central x_revs_x0000_added_x0000_copies effb563bb7e5 c07a39dc4e80 : 6 revs, 0.070508 s, 0.070294 s, -0.000214 s, ? 0.9970, 11715 ?s/rev
mozilla-central x000_revs_xx00_added_0_copies 6100d773079a 04a55431795e : 1593 revs, 0.006576 s, 0.006545 s, -0.000031 s, ? 0.9953, 4 ?s/rev
mozilla-central x000_revs_x000_added_x_copies 9f17a6fc04f9 2d37b966abed : 41 revs, 0.004809 s, 0.004998 s, +0.000189 s, ? 1.0393, 121 ?s/rev
mozilla-central x000_revs_x000_added_x000_copies 7c97034feb78 4407bd0c6330 : 7839 revs, 0.064872 s, 0.063348 s, -0.001524 s, ? 0.9765, 8 ?s/rev
mozilla-central x0000_revs_xx000_added_0_copies 9eec5917337d 67118cc6dcad : 615 revs, 0.026142 s, 0.026154 s, +0.000012 s, ? 1.0005, 42 ?s/rev
mozilla-central x0000_revs_xx000_added_x000_copies f78c615a656c 96a38b690156 : 30263 revs, 0.203956 s, 0.199063 s, -0.004893 s, ? 0.9760, 6 ?s/rev
mozilla-central x00000_revs_x0000_added_x0000_copies 6832ae71433c 4c222a1d9a00 : 153721 revs, 1.763853 s, 1.277320 s, -0.486533 s, ? 0.7242, 8 ?s/rev
mozilla-central x00000_revs_x00000_added_x000_copies 76caed42cf7c 1daa622bbe42 : 204976 revs, 2.609761 s, 1.698794 s, -0.910967 s, ? 0.6509, 8 ?s/rev
mozilla-try x_revs_x_added_0_copies aaf6dde0deb8 9790f499805a : 2 revs, 0.000847 s, 0.000842 s, -0.000005 s, ? 0.9941, 421 ?s/rev
mozilla-try x_revs_x000_added_0_copies d8d0222927b4 5bb8ce8c7450 : 2 revs, 0.000867 s, 0.000865 s, -0.000002 s, ? 0.9977, 432 ?s/rev
mozilla-try x_revs_x_added_x_copies 092fcca11bdb 936255a0384a : 4 revs, 0.000161 s, 0.000160 s, -0.000001 s, ? 0.9938, 40 ?s/rev
mozilla-try x_revs_x00_added_x_copies b53d2fadbdb5 017afae788ec : 2 revs, 0.001131 s, 0.001122 s, -0.000009 s, ? 0.9920, 561 ?s/rev
mozilla-try x_revs_x000_added_x000_copies 20408ad61ce5 6f0ee96e21ad : 1 revs, 0.033114 s, 0.032743 s, -0.000371 s, ? 0.9888, 32743 ?s/rev
mozilla-try x_revs_x0000_added_x0000_copies effb563bb7e5 c07a39dc4e80 : 6 revs, 0.071092 s, 0.071529 s, +0.000437 s, ? 1.0061, 11921 ?s/rev
mozilla-try x000_revs_xx00_added_0_copies 6100d773079a 04a55431795e : 1593 revs, 0.006554 s, 0.006593 s, +0.000039 s, ? 1.0060, 4 ?s/rev
mozilla-try x000_revs_x000_added_x_copies 9f17a6fc04f9 2d37b966abed : 41 revs, 0.005160 s, 0.005311 s, +0.000151 s, ? 1.0293, 129 ?s/rev
mozilla-try x000_revs_x000_added_x000_copies 1346fd0130e4 4c65cbdabc1f : 6657 revs, 0.065063 s, 0.063063 s, -0.002000 s, ? 0.9693, 9 ?s/rev
mozilla-try x0000_revs_x_added_0_copies 63519bfd42ee a36a2a865d92 : 40314 revs, 0.297118 s, 0.312363 s, +0.015245 s, ? 1.0513, 7 ?s/rev
mozilla-try x0000_revs_x_added_x_copies 9fe69ff0762d bcabf2a78927 : 38690 revs, 0.284002 s, 0.283106 s, -0.000896 s, ? 0.9968, 7 ?s/rev
mozilla-try x0000_revs_xx000_added_x_copies 156f6e2674f2 4d0f2c178e66 : 8598 revs, 0.086311 s, 0.083817 s, -0.002494 s, ? 0.9711, 9 ?s/rev
mozilla-try x0000_revs_xx000_added_0_copies 9eec5917337d 67118cc6dcad : 615 revs, 0.026738 s, 0.026516 s, -0.000222 s, ? 0.9917, 43 ?s/rev
mozilla-try x0000_revs_xx000_added_x000_copies 89294cd501d9 7ccb2fc7ccb5 : 97052 revs, 1.514270 s, 1.304865 s, -0.209405 s, ? 0.8617, 13 ?s/rev
mozilla-try x0000_revs_x0000_added_x0000_copies e928c65095ed e951f4ad123a : 52031 revs, 0.735875 s, 0.681088 s, -0.054787 s, ? 0.9255, 13 ?s/rev
mozilla-try x00000_revs_x_added_0_copies 6a320851d377 1ebb79acd503 : 363753 revs, 4.843329 s, 4.454320 s, -0.389009 s, ? 0.9197, 12 ?s/rev
mozilla-try x00000_revs_x00000_added_0_copies dc8a3ca7010e d16fde900c9c : 34414 revs, 0.591752 s, 0.567913 s, -0.023839 s, ? 0.9597, 16 ?s/rev
mozilla-try x00000_revs_x_added_x_copies 5173c4b6f97c 95d83ee7242d : 362229 revs, 4.760563 s, 4.547043 s, -0.213520 s, ? 0.9551, 12 ?s/rev
mozilla-try x00000_revs_x000_added_x_copies 9126823d0e9c ca82787bb23c : 359344 revs, 4.751942 s, 4.378579 s, -0.373363 s, ? 0.9214, 12 ?s/rev
mozilla-try x00000_revs_x0000_added_x0000_copies 8d3fafa80d4b eb884023b810 : 192665 revs, 2.605014 s, 1.703622 s, -0.901392 s, ? 0.6540, 8 ?s/rev
mozilla-try x00000_revs_x00000_added_x0000_copies 1b661134e2ca 1ae03d022d6d : 228985 revs, 41.113063 s, 36.001255 s, -5.111808 s, ? 0.8757, 157 ?s/rev
mozilla-try x00000_revs_x00000_added_x000_copies 9b2a99adc05e 8e29777b48e6 : 382065 revs, 27.891612 s, 14.340641 s, -13.550971 s, ? 0.5142, 37 ?s/rev
Differential Revision: https://phab.mercurial-scm.org/D9497
author | Pierre-Yves David <pierre-yves.david@octobus.net> |
---|---|
date | Sat, 21 Nov 2020 17:00:32 +0100 |
parents | b68b19104d16 |
children | e01e84e5e426 |
line wrap: on
line source
use byteorder::{BigEndian, ByteOrder}; /// A chunk of data to insert, delete or replace in a patch /// /// A chunk is: /// - an insertion when `!data.is_empty() && start == end` /// - an deletion when `data.is_empty() && start < end` /// - a replacement when `!data.is_empty() && start < end` /// - not doing anything when `data.is_empty() && start == end` #[derive(Debug, Clone)] struct Chunk<'a> { /// The start position of the chunk of data to replace start: u32, /// The end position of the chunk of data to replace (open end interval) end: u32, /// The data replacing the chunk data: &'a [u8], } impl Chunk<'_> { /// Adjusted start of the chunk to replace. /// /// The offset, taking into account the growth/shrinkage of data /// induced by previously applied chunks. fn start_offset_by(&self, offset: i32) -> u32 { let start = self.start as i32 + offset; assert!(start >= 0, "negative chunk start should never happen"); start as u32 } /// Adjusted end of the chunk to replace. /// /// The offset, taking into account the growth/shrinkage of data /// induced by previously applied chunks. fn end_offset_by(&self, offset: i32) -> u32 { self.start_offset_by(offset) + self.data.len() as u32 } /// Length of the replaced chunk. fn replaced_len(&self) -> u32 { self.end - self.start } /// Length difference between the replacing data and the replaced data. fn len_diff(&self) -> i32 { self.data.len() as i32 - self.replaced_len() as i32 } } /// The delta between two revisions data. #[derive(Debug, Clone)] pub struct PatchList<'a> { /// A collection of chunks to apply. /// /// Those chunks are: /// - ordered from the left-most replacement to the right-most replacement /// - non-overlapping, meaning that two chucks can not change the same /// chunk of the patched data chunks: Vec<Chunk<'a>>, } impl<'a> PatchList<'a> { /// Create a `PatchList` from bytes. pub fn new(data: &'a [u8]) -> Self { let mut chunks = vec![]; let mut data = data; while !data.is_empty() { let start = BigEndian::read_u32(&data[0..]); let end = BigEndian::read_u32(&data[4..]); let len = BigEndian::read_u32(&data[8..]); assert!(start <= end); chunks.push(Chunk { start, end, data: &data[12..12 + (len as usize)], }); data = &data[12 + (len as usize)..]; } PatchList { chunks } } /// Return the final length of data after patching /// given its initial length . fn size(&self, initial_size: i32) -> i32 { self.chunks .iter() .fold(initial_size, |acc, chunk| acc + chunk.len_diff()) } /// Apply the patch to some data. pub fn apply(&self, initial: &[u8]) -> Vec<u8> { let mut last: usize = 0; let mut vec = Vec::with_capacity(self.size(initial.len() as i32) as usize); for Chunk { start, end, data } in self.chunks.iter() { vec.extend(&initial[last..(*start as usize)]); vec.extend(data.iter()); last = *end as usize; } vec.extend(&initial[last..]); vec } /// Combine two patch lists into a single patch list. /// /// Applying consecutive patches can lead to waste of time and memory /// as the changes introduced by one patch can be overridden by the next. /// Combining patches optimizes the whole patching sequence. fn combine(&mut self, other: &mut Self) -> Self { let mut chunks = vec![]; // Keep track of each growth/shrinkage resulting from applying a chunk // in order to adjust the start/end of subsequent chunks. let mut offset = 0i32; // Keep track of the chunk of self.chunks to process. let mut pos = 0; // For each chunk of `other`, chunks of `self` are processed // until they start after the end of the current chunk. for Chunk { start, end, data } in other.chunks.iter() { // Add chunks of `self` that start before this chunk of `other` // without overlap. while pos < self.chunks.len() && self.chunks[pos].end_offset_by(offset) <= *start { let first = self.chunks[pos].clone(); offset += first.len_diff(); chunks.push(first); pos += 1; } // The current chunk of `self` starts before this chunk of `other` // with overlap. // The left-most part of data is added as an insertion chunk. // The right-most part data is kept in the chunk. if pos < self.chunks.len() && self.chunks[pos].start_offset_by(offset) < *start { let first = &mut self.chunks[pos]; let (data_left, data_right) = first.data.split_at( (*start - first.start_offset_by(offset)) as usize, ); let left = Chunk { start: first.start, end: first.start, data: data_left, }; first.data = data_right; offset += left.len_diff(); chunks.push(left); // There is no index incrementation because the right-most part // needs further examination. } // At this point remaining chunks of `self` starts after // the current chunk of `other`. // `start_offset` will be used to adjust the start of the current // chunk of `other`. // Offset tracking continues with `end_offset` to adjust the end // of the current chunk of `other`. let mut next_offset = offset; // Discard the chunks of `self` that are totally overridden // by the current chunk of `other` while pos < self.chunks.len() && self.chunks[pos].end_offset_by(next_offset) <= *end { let first = &self.chunks[pos]; next_offset += first.len_diff(); pos += 1; } // Truncate the left-most part of chunk of `self` that overlaps // the current chunk of `other`. if pos < self.chunks.len() && self.chunks[pos].start_offset_by(next_offset) < *end { let first = &mut self.chunks[pos]; let how_much_to_discard = *end - first.start_offset_by(next_offset); first.data = &first.data[(how_much_to_discard as usize)..]; next_offset += how_much_to_discard as i32; } // Add the chunk of `other` with adjusted position. chunks.push(Chunk { start: (*start as i32 - offset) as u32, end: (*end as i32 - next_offset) as u32, data, }); // Go back to normal offset tracking for the next `o` chunk offset = next_offset; } // Add remaining chunks of `self`. for elt in &self.chunks[pos..] { chunks.push(elt.clone()); } PatchList { chunks } } } /// Combine a list of patch list into a single patch optimized patch list. pub fn fold_patch_lists<'a>(lists: &[PatchList<'a>]) -> PatchList<'a> { if lists.len() <= 1 { if lists.is_empty() { PatchList { chunks: vec![] } } else { lists[0].clone() } } else { let (left, right) = lists.split_at(lists.len() / 2); let mut left_res = fold_patch_lists(left); let mut right_res = fold_patch_lists(right); left_res.combine(&mut right_res) } } #[cfg(test)] mod tests { use super::*; struct PatchDataBuilder { data: Vec<u8>, } impl PatchDataBuilder { pub fn new() -> Self { Self { data: vec![] } } pub fn replace( &mut self, start: usize, end: usize, data: &[u8], ) -> &mut Self { assert!(start <= end); self.data.extend(&(start as i32).to_be_bytes()); self.data.extend(&(end as i32).to_be_bytes()); self.data.extend(&(data.len() as i32).to_be_bytes()); self.data.extend(data.iter()); self } pub fn get(&mut self) -> &[u8] { &self.data } } #[test] fn test_ends_before() { let data = vec![0u8, 0u8, 0u8]; let mut patch1_data = PatchDataBuilder::new(); patch1_data.replace(0, 1, &[1, 2]); let mut patch1 = PatchList::new(patch1_data.get()); let mut patch2_data = PatchDataBuilder::new(); patch2_data.replace(2, 4, &[3, 4]); let mut patch2 = PatchList::new(patch2_data.get()); let patch = patch1.combine(&mut patch2); let result = patch.apply(&data); assert_eq!(result, vec![1u8, 2, 3, 4]); } #[test] fn test_starts_after() { let data = vec![0u8, 0u8, 0u8]; let mut patch1_data = PatchDataBuilder::new(); patch1_data.replace(2, 3, &[3]); let mut patch1 = PatchList::new(patch1_data.get()); let mut patch2_data = PatchDataBuilder::new(); patch2_data.replace(1, 2, &[1, 2]); let mut patch2 = PatchList::new(patch2_data.get()); let patch = patch1.combine(&mut patch2); let result = patch.apply(&data); assert_eq!(result, vec![0u8, 1, 2, 3]); } #[test] fn test_overridden() { let data = vec![0u8, 0, 0]; let mut patch1_data = PatchDataBuilder::new(); patch1_data.replace(1, 2, &[3, 4]); let mut patch1 = PatchList::new(patch1_data.get()); let mut patch2_data = PatchDataBuilder::new(); patch2_data.replace(1, 4, &[1, 2, 3]); let mut patch2 = PatchList::new(patch2_data.get()); let patch = patch1.combine(&mut patch2); let result = patch.apply(&data); assert_eq!(result, vec![0u8, 1, 2, 3]); } #[test] fn test_right_most_part_is_overridden() { let data = vec![0u8, 0, 0]; let mut patch1_data = PatchDataBuilder::new(); patch1_data.replace(0, 1, &[1, 3]); let mut patch1 = PatchList::new(patch1_data.get()); let mut patch2_data = PatchDataBuilder::new(); patch2_data.replace(1, 4, &[2, 3, 4]); let mut patch2 = PatchList::new(patch2_data.get()); let patch = patch1.combine(&mut patch2); let result = patch.apply(&data); assert_eq!(result, vec![1u8, 2, 3, 4]); } #[test] fn test_left_most_part_is_overridden() { let data = vec![0u8, 0, 0]; let mut patch1_data = PatchDataBuilder::new(); patch1_data.replace(1, 3, &[1, 3, 4]); let mut patch1 = PatchList::new(patch1_data.get()); let mut patch2_data = PatchDataBuilder::new(); patch2_data.replace(0, 2, &[1, 2]); let mut patch2 = PatchList::new(patch2_data.get()); let patch = patch1.combine(&mut patch2); let result = patch.apply(&data); assert_eq!(result, vec![1u8, 2, 3, 4]); } #[test] fn test_mid_is_overridden() { let data = vec![0u8, 0, 0]; let mut patch1_data = PatchDataBuilder::new(); patch1_data.replace(0, 3, &[1, 3, 3, 4]); let mut patch1 = PatchList::new(patch1_data.get()); let mut patch2_data = PatchDataBuilder::new(); patch2_data.replace(1, 3, &[2, 3]); let mut patch2 = PatchList::new(patch2_data.get()); let patch = patch1.combine(&mut patch2); let result = patch.apply(&data); assert_eq!(result, vec![1u8, 2, 3, 4]); } }