Which is more efficient w.r.t. CPU utilization between g++ and clang++? Let's find out with 30 input files and pay close attention to user times. g++ wins for parallel sort as lesser user time is better considering real times are similar.

cp big1.txt big4.txt
cp big2.txt big5.txt
cp big3.txt big6.txt

cp big1.txt big7.txt
cp big2.txt big8.txt
cp big3.txt big9.txt

cp big1.txt biga.txt
cp big2.txt bigb.txt
cp big3.txt bigc.txt

cp big1.txt bigd.txt
cp big2.txt bige.txt
cp big3.txt bigf.txt

cp big1.txt bigg.txt
cp big2.txt bigh.txt
cp big3.txt bigi.txt

cp big1.txt bigj.txt
cp big2.txt bigk.txt
cp big3.txt bigl.txt

cp big1.txt bigm.txt
cp big2.txt bign.txt
cp big3.txt bigo.txt

cp big1.txt bigp.txt
cp big2.txt bigq.txt
cp big3.txt bigr.txt

cp big1.txt bigs.txt
cp big2.txt bigt.txt
cp big3.txt bigu.txt
[download]

g++

$ g++ -o llil4vec-omp -std=c++20 -fopenmp -Wall -O3 llil4vec.cpp -I./f
+ast_io/include

$ time OMP_NUM_THREADS=30 ./llil4vec-omp big?.txt >out.txt

llil4vec (fixed string length=6) start
use OpenMP
get properties      time : 0.937948 secs
sort properties     time : 0.494592 secs
emplace set sort    time : 1.02668 secs
write stdout        time : 0.50569 secs
total               time : 2.96497 secs

real  0m3.271s   <-- total wall time, performs similarly
user  0m16.823s  <-- nothing unusual, fans seem quiet
sys   0m4.188s
[download]

clang++

$ clang++ -o llil4vec-omp -std=c++20 -fopenmp -Wall -O3 llil4vec.cpp -
+I./fast_io/include

$ time OMP_NUM_THREADS=30 ./llil4vec-omp big?.txt >out.txt

llil4vec (fixed string length=6) start
use OpenMP
get properties      time : 1.07873 secs
sort properties     time : 0.717261 secs
emplace set sort    time : 0.606499 secs
write stdout        time : 0.437287 secs
total               time : 2.83984 secs

real  0m3.201s   <-- total wall time, performs similarly
user  0m45.889s  <-- system fans noticeably louder
sys   0m5.934s
[download]

Next, I tried J Script with AVX2 enabled for comparison. 30 threads, specified in the script.

$ time ~/j904/bin/jconsole llil4_p.ijs big?.txt out.txt

Read and parse input:    2.43996
Classify, sum, sort:     4.63914
Format and write output: 0.848965
Total time:              7.92807

real  0m7.934s   <-- total wall time
user  0m24.782s  <-- this too, system fans roaring
sys   0m26.764s  <-- because user and sys combined
[download]

I find it interesting for clang++'s user time to be nearly 3x that of g++.