Rollup merge of rust-lang#61339 - jridgewell:pointer-alignment, r=BurntSushi

Optimize pointer alignment in utf8 validation

This uses (and reuses) the u8 arrays's inherent block alignment when checking whether the current index is block aligned.

I initially thought that this would just move the expensive `align_offset` call out of the while loop and replace it with a subtraction and bitwise AND. But it appears this optimizes much better, too...

before: https://rust.godbolt.org/z/WIPvWl
after: https://rust.godbolt.org/z/-jBPoW

## Benchmarks

https://github.com/jridgewell/faster-from_utf8/tree/pointer-alignment

```
test from_utf8_2_bytes_fast      ... bench:         310 ns/iter (+/- 42) = 1290 MB/s
test from_utf8_2_bytes_regular   ... bench:         309 ns/iter (+/- 24) = 1294 MB/s

test from_utf8_3_bytes_fast      ... bench:       1,027 ns/iter (+/- 62) = 1168 MB/s
test from_utf8_3_bytes_regular   ... bench:       1,513 ns/iter (+/- 611) = 793 MB/s

test from_utf8_4_bytes_fast      ... bench:       1,788 ns/iter (+/- 26) = 1342 MB/s
test from_utf8_4_bytes_regular   ... bench:       1,907 ns/iter (+/- 181) = 1258 MB/s

test from_utf8_all_bytes_fast    ... bench:       3,463 ns/iter (+/- 97) = 1155 MB/s
test from_utf8_all_bytes_regular ... bench:       4,083 ns/iter (+/- 89) = 979 MB/s

test from_utf8_ascii_fast        ... bench:          88 ns/iter (+/- 4) = 28988 MB/s
test from_utf8_ascii_regular     ... bench:          88 ns/iter (+/- 8) = 28988 MB/s

test from_utf8_cyr_fast          ... bench:       7,707 ns/iter (+/- 531) = 665 MB/s
test from_utf8_cyr_regular       ... bench:       8,202 ns/iter (+/- 135) = 625 MB/s

test from_utf8_enwik8_fast       ... bench:   1,135,756 ns/iter (+/- 84,450) = 8804 MB/s
test from_utf8_enwik8_regular    ... bench:   1,145,468 ns/iter (+/- 79,601) = 8730 MB/s

test from_utf8_jawik10_fast      ... bench:  12,723,844 ns/iter (+/- 473,247) = 785 MB/s
test from_utf8_jawik10_regular   ... bench:  13,384,596 ns/iter (+/- 666,997) = 747 MB/s

test from_utf8_mixed_fast        ... bench:       2,321 ns/iter (+/- 123) = 2081 MB/s
test from_utf8_mixed_regular     ... bench:       2,702 ns/iter (+/- 408) = 1788 MB/s

test from_utf8_mostlyasc_fast    ... bench:         249 ns/iter (+/- 10) = 14666 MB/s
test from_utf8_mostlyasc_regular ... bench:         276 ns/iter (+/- 5) = 13231 MB/s
```

Author: Centril

src/libcore/str/mod.rs

@@ -1422,6 +1422,7 @@ fn run_utf8_validation(v: &[u8]) -> Result<(), Utf8Error> {
     let usize_bytes = mem::size_of::<usize>();
     let ascii_block_size = 2 * usize_bytes;
     let blocks_end = if len >= ascii_block_size { len - ascii_block_size + 1 } else { 0 };
+    let align = v.as_ptr().align_offset(usize_bytes);
 
     while index < len {
         let old_offset = index;
@@ -1501,12 +1502,8 @@ fn run_utf8_validation(v: &[u8]) -> Result<(), Utf8Error> {
             // Ascii case, try to skip forward quickly.
             // When the pointer is aligned, read 2 words of data per iteration
             // until we find a word containing a non-ascii byte.
-            let ptr = v.as_ptr();
-            let align = unsafe {
-                // the offset is safe, because `index` is guaranteed inbounds
-                ptr.add(index).align_offset(usize_bytes)
-            };
-            if align == 0 {
+            if align.wrapping_sub(index) % usize_bytes == 0 {
+                let ptr = v.as_ptr();
                 while index < blocks_end {
                     unsafe {
                         let block = ptr.add(index) as *const usize;