core: Fix overflow in int::mod_euc when self < 0 && rhs == MIN

[tesaguri]

This commit removes usage of abs, which overflows when self == MIN.

[rust-highfive]

r? @Kimundi

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

[varkor]

This line and the previous might as well be consistent (whichever of < 0 or is_negative you prefer).

[varkor]

This was my fault for not paying close enough attention during review! Thanks for catching it!

[kennytm]

@bors r+ rollup

[bors]

📌 Commit 104c64d has been approved by kennytm

[bors]

☔ The latest upstream changes (presumably #50191) made this pull request unmergeable. Please resolve the merge conflicts.

[Boscop]

I'd prefer mod_euc to have no branches, because it often occurs in very hot loops in my code.
Can we change it to remove the branches? So instead of this:

pubfnmod_euc(self,rhs:Self) -> Self{let r = self % rhs;if r < 0{if rhs < 0{
r - rhs
}else{
r + rhs
}}else{
r
}}

Something like this:

pubfnmod_euc(self,rhs:Self) -> Self{let r = self % rhs;let r_pos = [r + rhs, r - rhs][(rhs < 0)asusize];[r, r_pos][(r < 0)asusize)]}

The compiler should be able to remove the array bounds checks, right?

[varkor]

@Boscop: I'm curious to know how differently this performs — have you written a benchmark to see? (Also, maybe it would be better to continue this on the tracking issue where it's more discoverable!)

[Boscop]

@varkor
Nvm, it seems that the current version is the fastest (tested on x64 Win 8.1 with i64 ints only):

test tests::bench_mod_euc_branching ... bench: 33,072,518 ns/iter (+/- 1,355,002)
test tests::bench_mod_euc_branchless ... bench: 34,965,845 ns/iter (+/- 1,265,057)
test tests::bench_mod_euc_branchless_2 ... bench: 35,493,871 ns/iter (+/- 838,259)
test tests::bench_mod_euc_branchless_u ... bench: 35,033,348 ns/iter (+/- 1,717,562)

Is this the correct way to write a benchmark?

#[cfg(test)]mod tests {use test::{Bencher, black_box};pubfnmod_euc_branching(slf:i64,rhs:i64) -> i64{let r = slf % rhs;if r < 0{if rhs < 0{
r - rhs
}else{
r + rhs
}}else{
r
}}pubfnmod_euc_branchless(slf:i64,rhs:i64) -> i64{let r = slf % rhs;let r_pos = [r + rhs, r - rhs][(rhs < 0)asusize];[r, r_pos][(r < 0)asusize]}pubfnmod_euc_branchless_u(slf:i64,rhs:i64) -> i64{let r = slf % rhs;unsafe{let r_pos = *[r + rhs, r - rhs].get_unchecked((rhs < 0)asusize);*[r, r_pos].get_unchecked((r < 0)asusize)}}pubfnmod_euc_branchless_2(slf:i64,rhs:i64) -> i64{let r = slf % rhs;let r_pos = r + rhs - 2* rhs *(rhs < 0)asi64;
r + (r_pos - r)*(r < 0)asi64}#[bench]fnbench_mod_euc_branching(b:&mutBencher){
b.iter(|| {for a in -1000..1000{for b in -1000..1000{if b == 0{continue;}black_box(mod_euc_branching(a, b));}}});}#[bench]fnbench_mod_euc_branchless(b:&mutBencher){
b.iter(|| {for a in -1000..1000{for b in -1000..1000{if b == 0{continue;}black_box(mod_euc_branchless(a, b));}}});}#[bench]fnbench_mod_euc_branchless_u(b:&mutBencher){
b.iter(|| {for a in -1000..1000{for b in -1000..1000{if b == 0{continue;}black_box(mod_euc_branchless_u(a, b));}}});}#[bench]fnbench_mod_euc_branchless_2(b:&mutBencher){
b.iter(|| {for a in -1000..1000{for b in -1000..1000{if b == 0{continue;}black_box(mod_euc_branchless_2(a, b));}}});}}

[varkor]

@Boscop: yeah, that looks fine to me. Thanks for checking it out!