Missed optimization: codegen test repeat-trusted-len.rs fails when stdlib is compiled with overflow-checks=on

[alex]

This is another discovery in my work on #58475 / rust-lang/rfcs#2635 -- trying to see if it's practical to enable overflow checks by default in release builds.

This fails with:

stderr:
------------------------------------------
/home/alex/p/rust/src/test/codegen/repeat-trusted-len.rs:11:11: error: CHECK: expected string not found in input
// CHECK: call void @llvm.memset.p0i8.i{{[0-9]+}}(i8* {{(nonnull )?}}align 1{{.*}} %{{[0-9]+}}, i8 42, i{{[0-9]+}} 100000, i1 false)
          ^
/home/alex/p/rust/build/x86_64-unknown-linux-gnu/test/codegen/repeat-trusted-len/repeat-trusted-len.ll:51:33: note: scanning from here
define void @repeat_take_collect(%"alloc::vec::Vec<u8>"* noalias nocapture sret dereferenceable(24) %0) unnamed_addr #2 personality i32 (i32, i32, i64, %"unwind::libunwind::_Unwind_Exception"*, %"unwind::libunwind::_Unwind_Context"*)* @rust_eh_personality {

The issue seems to be that the LLVM loop-idiom recognizer runs before the optimizations that remove the overflow checks, therefor you get a loop in assembly, instead of a call to memset. I've filed a bug against LLVM with this: https://bugs.llvm.org/show_bug.cgi?id=46057 , but I'm also filing a bug here so there's a way to track the rust side impact of it

[the8472]

Odd, this doesn't reproduce on godbolt (i.e. there's a memset), even when I go back to a release from 2020.

https://rust.godbolt.org/z/WjM7jz619

[alex]

That's using a stdlib that's not compiled with overflow-checks -- the overflow checks only applies to your code.

…

On Mon, Oct 11, 2021 at 3:11 PM the8472 ***@***.***> wrote: Odd, this doesn't reproduce on godbolt (i.e. there's a memset), even when I go back to a release from 2020. https://rust.godbolt.org/z/WjM7jz619 — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#72549 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAAAGBCMNGY3BZOBJIVGTQLUGMZG5ANCNFSM4NJBYZVQ> . Triage notifications on the go with GitHub Mobile for iOS <https://apps.apple.com/app/apple-store/id1477376905?ct=notification-email&mt=8&pt=524675> or Android <https://play.google.com/store/apps/details?id=com.github.android&referrer=utm_campaign%3Dnotification-email%26utm_medium%3Demail%26utm_source%3Dgithub>.

-- All that is necessary for evil to succeed is for good people to do nothing.

[the8472]

Right, I forgot, although it depends on whether the methods in question are annotated with #[rustc_inherit_overflow_checks].

[the8472]

We could work around that LLVM shortcoming by using unchecked add/sub which have been added in the meantime. #85122

[hkratz]

Can't seem to reproduce it. Maybe it is gone with the new LLVM pass manager?

[alex]

Maybe? The C reproucer I filed the LLVM bug with still reproduces for me.

[alex]

The LLVM bug I filed no longer reproduces, so I'm going to close this.