[llvm-bugs] [Bug 52455] New: Repeated conditionals are not combined with differing conditionals in between

via llvm-bugs llvm-bugs at lists.llvm.org
Tue Nov 9 12:34:35 PST 2021


            Bug ID: 52455
           Summary: Repeated conditionals are not combined with differing
                    conditionals in between
           Product: libraries
           Version: 13.0
          Hardware: PC
                OS: Linux
            Status: NEW
          Severity: enhancement
          Priority: P
         Component: Scalar Optimizations
          Assignee: unassignedbugs at nondot.org
          Reporter: llvm-bugs at bluematt.me
                CC: llvm-bugs at lists.llvm.org

In some auto-generated XDP code which (in part) consists of a long list of
conditional -> return XDP_DROP statements, LLVM 13 seems materially worse at
combining parts of the conditionals which are repeated than LLVM 12-, though
both miss some major cases.

The godbolt link at https://godbolt.org/z/hW3r6EKqT demonstrates this using
something that is based on my (real) auto-generated code. In clang 13, the
first part of the repeated conditional is identical and combined, but any parts
after the `ip->saddr` check (which differs across conditionals) is repeated,
almost doubling the number of instructions the BPF verifier has to process.

In clang 12 (and most previous versions), LLVM can sometimes remove the
repeated `dport` check, but only if its the only if `SADDRA_MASK` and
`SADDRB_MASK` are equal or EXTRA_CHECK is empty. In the real auto-generated
code, `SADDRA_MASK` and `SADDRB_MASK` differ across each conditional, and
`EXTRA_CHECK` is something like "ip->protocol == 6 &&", which causes clang
11/12 to repeat the port and protocol checks for each conditional (ie
https://godbolt.org/z/h8d1zesjr which is the same base code as above).

You are receiving this mail because:
You are on the CC list for the bug.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-bugs/attachments/20211109/bb24c0f8/attachment.html>

More information about the llvm-bugs mailing list