user0/rust
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Remove c32() from x86_64 memcmp

Browse source 
Fixes https://github.com/rust-lang/compiler-builtins/issues/487
This commit is contained in:
David Hoppenbrouwers 2022-08-10 11:29:38 +02:00
parent 09aca6f001
commit a695cf95cf
No known key found for this signature in database
GPG key ID: A9156EA5E4B644FF
1 changed files with 1 additions and 13 deletions
Download patch file Download diff file Expand all files Collapse all files

14
library/compiler-builtins/src/mem/x86_64.rs
View file

@ -170,19 +170,7 @@ pub unsafe fn compare_bytes(a: *const u8, b: *const u8, n: usize) -> i32 {
let c4 = |a: *const u32, b, n| cmp(a, b, n, c2);
let c8 = |a: *const u64, b, n| cmp(a, b, n, c4);
let c16 = |a: *const u128, b, n| cmp(a, b, n, c8);
let c32 = |a: *const [u128; 2], b, n| cmp(a, b, n, c16);
// [u128; 2] internally uses raw_eq for comparisons, which may emit a call to memcmp
// above a certain size threshold. When SSE2 is enabled this threshold does not seem
// to be reached but without SSE2 a call is emitted, leading to infinite recursion.
//
// While replacing [u128; 2] with (u128, u128) fixes the issues it degrades performance
// severely. Likewise, removing c32() has a lesser but still significant impact. Instead the
// [u128; 2] case is only enabled when SSE2 is present.
if cfg!(target_feature = "sse2") {
c32(a.cast(), b.cast(), n)
} else {
c16(a.cast(), b.cast(), n)
}
c16(a.cast(), b.cast(), n)
}
/// Determine optimal parameters for a `rep` instruction.