[AArch64] Extremely slow code generation for series of function calls/addition

[llvmbot]

Bugzilla Link	51147
Version	trunk
OS	Linux
Attachments	.c file producing failure
Reporter	LLVM Bugzilla Contributor

Extended Description

In a function containing about 1000 calls and adds, we see the aarch64 target taking about 30 minutes to do code generation. The slow pass is:

BISECT: running pass (96) AArch64 Instruction Selection on function (len_internal)

Other targets have no issue and complete quickly.

gcc time: 0.617u 0.022s 0:00.65 96.9% 0+0k 32+160io 0pf+0w
clang aarch64 time: 1998.064u 0.135s 33:23.82 99.7% 0+0k 0+256io 0pf+0w
clang time default target: 0.706u 0.020s 0:00.74 97.2% 0+0k 0+144io 0pf+0w

Command that is slow: clang --target=aarch64-unknown-linux -O3 -c simple.c

Clang version - latest from trunk:
clang version 13.0.0 (https://github.com/llvm/llvm-project 9da70ab)

Attached the file that causes the slowdown

[vielmetti]

Report from https://twitter.com/p_paralkar/status/1525230176107483138 p_paralkar:

perf(1) tells us: llc -march=aarch64 -O3 simple.bc spends:
       78.96%  llc      llc                  [.] (anonymous namespace)::popFromQueueImpl<(anonymous namespace)::src_ls_rr_sort>

(simple.bc is the LLVM IR from the given test case).

[llvmbot]

@llvm/issue-subscribers-backend-aarch64

[anmol-paralkar]

Analysis: analysis.txt

[anmol-paralkar]

With -mllvm --combiner-select-seq at -O3 on:

Linux llvm-arm 5.4.0-92-generic #103-Ubuntu SMP Fri Nov 26 16:15:10 UTC 2021 aarch64 aarch64 aarch64 GNU/Linux

using the above test case converted into a little benchmark:
test-1100.tar.gz

we get:

-rwxr-xr-x 1 root root 129992 Oct 17 01:04 test-1100.pv.O3.64.exe
-rwxr-xr-x 1 root root 109856 Oct 17 01:04 test-1100.css.O3.64.exe

• about 15.4902% space savings, and:

root@llvm-arm:~/test-meta# ./driver.sh 1100 [4/1802]
Compiling 64 bit at -O3
Compiling baseline:

real 262m6.496s
user 261m57.975s
sys 0m0.148s
Compiling test:

real 0m8.461s
user 0m8.381s
sys 0m0.076s
Running 64 bit at -O3 pv
iterations: 100000000

real 7m21.891s
user 7m21.829s
sys 0m0.000s
Running 64 bit at -O3 css
iterations: 100000000

real 4m54.228s
user 4m54.193s
sys 0m0.000s
root@llvm-arm:~/test-meta#

i.e. 99.9462% improvement in compile time and 33.4162% improvement in runtime.

Thanks to the Works On Arm program https://www.arm.com/solutions/infrastructure/works-on-arm for providing the equinix-metal-arm64-cluster by Equinix https://www.equinix.com/ for developing this solution.

[anmol-paralkar]

Test cases that characterize the space around the issue:
test-meta.tar.gz