[llvm-bugs] [Bug 52146] New: [SystemZ][NewPM] huge loop expansion with opt -O1

[via llvm-bugs]

https://bugs.llvm.org/show_bug.cgi?id=52146

            Bug ID: 52146
           Summary: [SystemZ][NewPM] huge loop expansion with opt -O1
           Product: libraries
           Version: trunk
          Hardware: PC
                OS: Linux
            Status: NEW
          Severity: normal
          Priority: P
         Component: Backend: SystemZ
          Assignee: unassignedbugs at nondot.org
          Reporter: jistone at redhat.com
                CC: llvm-bugs at lists.llvm.org

Created attachment 25352
  --> https://bugs.llvm.org/attachment.cgi?id=25352&action=edit
bitcode from rust#89609

xref: https://github.com/rust-lang/rust/issues/89609

After enabling the new pass manager in rustc, one particular crate showed a
large increase in compile time, but only when targeting s390x. I attached the
bitcode from rustc, and I can reproduce the problem with opt -O1 on main.

$ llvm-dis <rustc_ast_lowering-cgu.0.rcgu.thin-lto-after-patch.bc | wc -lL
 157863    2306

$ time opt <rustc_ast_lowering-cgu.0.rcgu.thin-lto-after-patch.bc -O1 -S | wc
-lL
 794414 4515800

real    2m4.580s
user    2m4.675s
sys     0m0.221s

There are two instances of that longest 4.5M line length, and both are huge phi
nodes just before invoking a function ending in 17hde8a472161ebd31bE (so you
can search for that). The predecessors look like a mass of loopexit.split-lp
expansions gone rogue.

The same input behaves reasonably with the old PM, both optimizing faster and
resulting in a smaller bitcode.

$ time opt <rustc_ast_lowering-cgu.0.rcgu.thin-lto-after-patch.bc -O1 -S
--enable-new-pm=0 | wc -lL
 148220    2306

real    0m2.266s
user    0m2.322s
sys     0m0.024s


If you edit out the "target-cpu"="z10" attributes in the IR, this reproducer
can also be forced to -mtriple x86_64, and that runs fine with new PM.

-- 
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/20211012/0bcb80e9/attachment.html>